Update `git_win32_path_remove_namespace` to disambiguate the prefix
being removed versus the prefix being added.  Now we remove the
"namespace", and (may) add a "prefix" in its place.  Eg, we remove the
`\\?\` namespace.  We remove the `\\?\UNC\` namespace, and replace it
with the `\\` prefix.  This aids readability somewhat.

Additionally, use pointer arithmetic instead of offsets, which seems to
also help readability.

The internal API `git_win32__canonicalize_path` is far, far too easily
confused with the internal API `git_win32_path_canonicalize`.  The
former removes the namespace prefix from a path (eg, given
`\\?\C:\Temp\foo`, it returns `C:\Temp\foo`, and given
`\\?\UNC\server\share`, it returns `\\server\share`).  As such, rename
it to `git_win32_path_remove_namespace`.

`git_win32_path_canonicalize` remains unchanged.

Now that we have abstracted away how to store and retrieve config
entries, it became trivial to implement a new in-memory backend by
making use of this. And thus we do so.

This commit implements a new read-only in-memory backend that can parse
a chunk of memory into a `git_config_backend` structure.

Right now, the config file code requires us to pass in its backend to
the config entry iterator. This is required with the current code, as
the config file backend will first create a read-only snapshot which is
then passed to the iterator just for that purpose. So after the iterator
is getting free'd, the code needs to make sure that the snapshot gets
free'd, as well.

By now, though, we can easily refactor the code to be more efficient and
remove the reverse dependency from iterator to backend. Instead of
creating a read-only snapshot (which also requires us to re-parse the
complete configuration file), we can simply duplicate the config entries
and pass those to the iterator. Like that, the iterator only needs to
make sure to free the duplicated config entries, which is trivial to do
and clears up memory ownership by a lot.

Access to the config entries is now completely done via the modules
function interface and no caller messes with the struct's internals. We
can thus completely move the structure declarations into the
implementation file so that nobody even has a chance to mess with the
members.

Instead of directly calling `git_atomic_inc` in users of the config
entries store, provide a `git_config_entries_incref` function to further
decouple the interfaces. Convert the refcount to a `git_refcount`
structure while at it.

The nice thing about our `git_config_iterator` interfaces is that nobody
needs to know anything about the implementation details. All that is
required is to obtain the iterator via any backend and then use it by
executing generic functions. We can thus completely internalize all the
implementation details of how to iterate over entries into the config
entries store and simply create such an iterator in our config file
backend when we want to iterate its entries. This further decouples the
config file backend from the config entries store.

The code accessing config entries in the `git_config_entries` structure
is still much too intimate with implementation details, directly
accessing the maps and handling indices. Provide two new functions to
get config entries from the internal map structure to decouple the
interfaces and use them in the config file code.

The function `git_config_entries_get` will simply look up the entry by
name and, in the case of a multi-value, return the last occurrence of
that entry. The second function, `git_config_entries_get_unique`, will
only return an entry if it is unique and not included via another
configuration file. This one is required to properly implement write
operations for single entries, as we refuse to write to or delete a
single entry if it is not clear which one was meant.

The previous commit simply moved all code that is required to handle
config entries to a new module without yet adjusting any of the function
and structure names to help readability. We now rename things
accordingly to have a common "git_config_entries" entries instead of the
old "diskfile_entries" one.

The configuration entry store that is used for configuration files needs
to keep track of all entries in two different structures:

- a singly linked list is being used to be able to iterate through
  configuration files in the order they have been found

- a string map is being used to efficiently look up configuration
  entries by their key

This store is thus something that may be used by other, future backends
as well to abstract away implementation details and iteration over the
entries.

Pull out the necessary functions from "config_file.c" and moves them
into their own "config_entries.c" module. For now, this is simply moving
over code without any renames and/or refactorings to help reviewing.

The implementation for config file snapshots has an unnecessary
redirection from `config_snapshot` to `git_config_file__snapshot`.
Inline the call to `git_config_file__snapshot` and remove it.

The header "config_file.h" has a list of inline-functions to access the
contents of a config backend without directly messing with the struct's
function pointers. While all these functions are called
"git_config_file_*", they are in fact completely backend-agnostic and
don't care whether it is a file or not. Rename all the function to
instead be backend-agnostic versions called "git_config_backend_*" and
rename the header to match.

The function `git_config_file_normalize_section` is never being used in
any file different than "config.c", but it is implemented in
"config_file.c". Move it over and make the symbol static.

As a last step to make variables and structures more backend agnostic
for our `git_config` structure, rename local variables to not be called
`file` anymore.

"warning: values of type 'OSStatus' should not be used as format arguments; add an explicit cast to 'int' instead [-Wformat]"

Same as with the previous commit, the `file_internal` struct is used to
keep track of all the backends that are added to a `git_config` struct.
Rename it to `backend_internal` and rename its `file` member to
`backend` to make the implementation more backend-agnostic.

Originally, the `git_config` struct is a collection of all the parsed
configuration files from different scopes (system-wide config,
user-specific config as well as the repo-specific config files).
Historically, we didn't and don't yet have any other configuration
backends than the one for files, which is why the field holding the
config backends is called `files`. But in fact, nothing dictates that
the vector of backends actually holds file backends only, as they are
generic and custom backends can be implemented by users.

Rename the member to be called `backends` to clarify that there is
nothing specific to files here.

The variables `git_config_escaped` and `git_config_escapes` are both
defined as static const character pointers in "config_parse.h". In case
where "config_parse.h" is included but those two variables are not being
used, the compiler will thus complain about defined but unused
variables. Fix this by declaring them as external and moving the actual
initialization to the C file.

Note that it is not possible to simply make this a #define, as we are
indexing into those arrays.

When populating the list of submodule names, we use the submodule
configuration entry's name as the key in the map of submodule names.
This creates a hidden dependency on the liveliness of the configuration
that was used to parse the submodule, which is fragile and unexpected.

Fix the issue by duplicating the string before writing it into the
submodule name map.

This is not a big deal, but it does make us match git more closely by checking
only the first. The lists are sorted already, so there should be no functional
difference other than removing a possible check from every iteration in the
loop.

When porting, we overlooked that the difference between git's and our's time
representation and copied their way of getting the max value.

Unfortunately git was using unsigned integers, so `~0ll` does correspond to
their max value, whereas for us it corresponds to `-1`. This means that we
always consider the last date to be smaller than the current commit's and always
think commits are interesting.

Change the initial value to the macro that gives us the maximum value on each
platform so we can accurately consider commits interesting or not.

ARM treats its `char` type as `unsigned type` by default; as a result,
testing a `char` value as being `< 0` is always false.  This is a
warning on ARM, which is promoted to an error given our use of
`-Werror`.

Per ISO 9899:199, section "6.2.5 Types":

> The three types char, signed char, and unsigned char are collectively
> called the character types. The implementation shall define char to
> have the same range, representation, and behavior as either signed
> char or unsigned char.
>
...

> Irrespective of the choice made, char is a separate type from the other
> two and is not compatible with either.

Previously we would assert in index_free because the reader incrementation
would not be balanced. Release the snapshot normally, so the variable gets
decremented before the index is freed.

This makes it easier to pass connection-related options around (proxy &
custom headers for now).

This fixes a bug in git_push_finish, which didn't reuse the provided
proxy if the connection closed between the call to `git_remote_push` and
the finish step.

In case where we add multiple configuration entries with the same key to
a diskfile backend, we always need to iterate the list of this key to
find the last entry due to the list being a singly-linked list. This
is obviously quadratic behaviour, and this has sure enough been found by
oss-fuzz by generating a configuration file with 50k lines, where most
of them have the same key. While the issue will not arise with "sane"
configuration files, an adversary may trigger it by providing a crafted
".gitmodules" file, which is delivered as part of the repo and also
parsed by the configuration parser.

The fix is trivial: store a pointer to the last entry of the list in its
head. As there are only two locations now where we append to this data
structure, mainting this pointer is trivial, too. We can also optimize
retrieval of a single value via `config_get`, where we previously had to
chase the `next` pointer to find the last entry that was added.

Using our configuration file fozzur with a corpus that has a single file
with 50000 "-=" lines previously took around 21s. With this optimization
the same file scans in about 0.053s, which is a nearly 400-fold
improvement. But in most cases with a "normal" amount of same-named keys
it's not going to matter anyway.

When running repack while doing repo writes, `packfile_load__cb()` can see some temporary files in the directory that are bigger than the usual, and makes `memcmp` overflow on the `p->pack_name` string. ASAN detected this. This just uses `strncmp`, that should not have any performance impact and is safe for comparing strings of different sizes.

```
ERROR: AddressSanitizer: heap-buffer-overflow on address 0x61200001a3f3 at pc 0x7f4a9e1976ec bp 0x7ffc1f80e100 sp 0x7ffc1f80d8b0
READ of size 89 at 0x61200001a3f3 thread T0
SCARINESS: 26 (multi-byte-read-heap-buffer-overflow)
    #0 0x7f4a9e1976eb in __interceptor_memcmp.part.78 (/build/cfgr-admin#link-tree/libtools_build_sanitizers_asan-ubsan-py.so+0xcf6eb)
    #1 0x7f4a518c5431 in packfile_load__cb /build/libgit2/0.27.0/src/libgit2-0.27.0/src/odb_pack.c:213
    #2 0x7f4a518d9582 in git_path_direach /build/libgit2/0.27.0/src/libgit2-0.27.0/src/path.c:1134
    #3 0x7f4a518c58ad in pack_backend__refresh /build/libgit2/0.27.0/src/libgit2-0.27.0/src/odb_pack.c:347
    #4 0x7f4a518c1b12 in git_odb_refresh /build/libgit2/0.27.0/src/libgit2-0.27.0/src/odb.c:1511
    #5 0x7f4a518bff5f in git_odb__freshen /build/libgit2/0.27.0/src/libgit2-0.27.0/src/odb.c:752
    #6 0x7f4a518c17d4 in git_odb_stream_finalize_write /build/libgit2/0.27.0/src/libgit2-0.27.0/src/odb.c:1415
    #7 0x7f4a51b9d015 in Repository_write /build/pygit2/0.27.0/src/pygit2-0.27.0/src/repository.c:509
```

The current error handling for the multiline variable parser is a bit
fragile, as each error condition has its own code to clear memory.
Instead, unify error handling as far as possible to avoid this
repetitive code. While at it, make use of `GITERR_CHECK_ALLOC` to
correctly handle OOM situations and verify that the buffer we print into
does not run out of memory either.

This performs a compile-check by using CMake support, to differentiate the GNU
version from the BSD version of qsort_r.

Module taken from 4f252abea5f1d17c60f6ff115c9c44cc0b6f1df6, which I've checked
against CMake 2.8.11.

Otherwise we return a NULL context, which will get dereferenced in 
apply_credentials.

Otherwise we'll return stack data to the caller.

In case there was nothing to parse in the buf, we'd return uninitialized
stack data.

At line 594, we do this :
if (error < 0)
		return error;

but if nothing was pushed in a GIT_SORT_TIME revwalk, we'd return
uninitialized stack data.