Cache the parsed submodule config when diffing

Make git__strntol64() ~70%* faster

winhttp: support optional client cert

This change uses compiler intrinsics to detect overflows instead of
using divisions to detect potential overflow. This makes the code faster
and makes it easier to read as a bonus side-effect!

Some of the things this quickens:

* Config parsing.
* Tree parsing.
* Smart protocol negotiation.

\* Measured by running `libgit2_clar` with `-fno-optimize-sibling-calls
-fno-omit-frame-pointer` under `perf(1)`:

```shell
$ perf diff --symbols=git__strntol64 --compute=ratio \
  --percentage=absolute baseline.data perf.data
\# Event 'cycles'
\#
\# Baseline           Ratio  Shared Object
\# ........  ..............  .............
\#
     0.25%        0.321836  libgit2_clar
```

Also repurposed an unused function and deleted another one.

This change makes that anything that calls `git_diff__from_iterators`
(any of the `git_diff_xxx` functions) only need to parse the
`.gitmodules` file once. This can be avoided by calling
`git_repository_submodule_cache_all(...)`, but we can do that safely for
the user with no change in semantics.

Fixes: #5725

git.git-authors: Replacing his/her with their

Friendlier getting started in the lack of git_libgit2_init

We require the library to be initialized with git_libgit2_init before it
is functional.  However, if a user tries to uses the library without
doing so - as they might when getting started with the library for the
first time - we will likely crash.

This commit introduces some guard rails - now instead of having _no_
allocator by default, we'll have an allocator that always fails, and
never tries to set an error message (since the thread-local state is
set up by git_libgit2_init).  We've modified the error retrieval
function to (try to) ensure that the library has been initialized before
getting the thread-local error message.

(Unfortunately, we cannot determine if the thread local storage has
actually been configured, this does require initialization by
git_libgit2_init.  But a naive attempt should be good enough for most
cases.)

A function to provide the initialization count of the library; this is
subject to race conditions but is useful for a naive determination as to
whether the library has been initialized or not.

Thread-local storage: a generic internal library (with no allocations)

Use a no-allocation approach to the TLS data abstraction.

Provide more clarity for Win32 calling conventions - now that we always
compile to __cdecl on Win32, we'll define that as the the libgit2
calling convention.  Also offer NTAPI (__stdcall) calling conventions
for things that need callbacks from NTAPI code (eg fiber-local storage).

This adds a generic TLS interface for anyone to store TLS data. It is designed to work regardless of whether threading support is built into the library or not.

Nobody in the library should directly interface with the data on the TLS struct, so it's been built to be opaque even in the library.

Requires the allocator to be initialized before use.

Clarify the `git_atomic` type and functions now that we have a 64 bit
version as well (`git_atomic64`).

The number of CPUs is useful information for creating a thread pool or a
number of workers, but it's not really about threading directly.  Evict
it from the thread file

Thread-free implementation

Our git_packbuilder__cache_lock function returns a value; use git__noop.

The git__noop function is more largely useful; move it into the
util header.  (And reduce the number of underscores.)

Make the pack and mwindow implementations data-race-free

Make the odb race-free

This change fixes a packfile heap corruption that can happen when
interacting with multiple packfiles concurrently across multiple
threads. This is exacerbated by setting a lower mwindow open file limit.

This change:

* Renames most of the internal methods in pack.c to clearly indicate
  that they expect to be called with a certain lock held, making
  reasoning about the state of locks a bit easier.
* Splits the `git_pack_file` lock in two: the one in `git_pack_file`
  only protects the `index_map`. The protection to `git_mwindow_file` is
  now in that struct.
* Explicitly checks for freshness of the `git_pack_file` in
  `git_packfile_unpack_header`: this allows the mwindow implementation
  to close files whenever there is enough cache pressure, and
  `git_packfile_unpack_header` will reopen the packfile if needed.
* After a call to `p_munmap()`, the `data` and `len` fields are poisoned
  with `NULL` to make use-after-frees more evident and crash rather than
  being open to the possibility of heap corruption.
* Adds a test case to prevent this from regressing in the future.

Fixes: #5591

This change adds all the necessary locking to the odb to avoid races in
the backends.

Part of: #5592

Also add the raw hostkey to `git_cert_hostkey`

Fix the `ENABLE_WERROR=ON` build in Groovy Gorilla (gcc 10.2)

The release build has been failing with the following error:

```shell
FAILED: src/CMakeFiles/git2internal.dir/hash.c.o
/usr/bin/cc -DHAVE_QSORT_R_GNU -DSHA1DC_CUSTOM_INCLUDE_SHA1_C=\"common.h\" -DSHA1DC_CUSTOM_INCLUDE_UBC_CHECK_C=\"common.h\" -DSHA1DC_NO_STANDARD_INCLUDES=1 -D_FILE_OFFSET_BITS=64 -Isrc -I../src -I../include -I../deps/pcre -I../deps/http-parser -D_GNU_SOURCE -fsanitize=thread -fno-optimize-sibling-calls -fno-omit-frame-pointer -Werror -Wall -Wextra -fvisibility=
hidden -fPIC -Wno-documentation-deprecated-sync -Wno-missing-field-initializers -Wstrict-aliasing -Wstrict-prototypes -Wdeclaration-after-statement -Wshift-count-overflow -Wunused-const-variable -Wunused-function -Wint-conversion -Wformat -Wformat-security -Wmissing-declarations -g -D_DEBUG -O0   -std=gnu90 -MD -MT src/CMakeFiles/git2internal.dir/hash.c.o -MF s
rc/CMakeFiles/git2internal.dir/hash.c.o.d -o src/CMakeFiles/git2internal.dir/hash.c.o   -c ../src/hash.c
../src/hash.c: In function ‘git_hash_init’:
../src/hash.c:47:1: error: control reaches end of non-void function [-Werror=return-type]
   47 | }
      | ^
../src/hash.c: In function ‘git_hash_update’:
../src/hash.c:58:1: error: control reaches end of non-void function [-Werror=return-type]
   58 | }
      | ^
../src/hash.c: In function ‘git_hash_final’:
../src/hash.c:68:1: error: control reaches end of non-void function [-Werror=return-type]
   68 | }
      | ^
../src/hash.c: At top level:
cc1: note: unrecognized command-line option ‘-Wno-documentation-deprecated-sync’ may have been intended to silence earlier diagnostics
cc1: all warnings being treated as errors
[11/533] Building C object src/CMakeFiles/git2internal.dir/odb_pack.c.o
ninja: build stopped: subcommand failed.
```

The compiler _should_ be able to figure out that there is no way to reach the
end of the non-void function since `GIT_ASSERT(0)` expands to either `assert()`
or an unconditional `return -1;` (after doing constant folding and stuff,
depending on the debug level). But it's not doing so at the moment, so let's
help it.

`git_cert_x509` has the raw encoded certificate. Let's do the same for
the SSH certificate for symmetry.

This change adds support for reading multi-pack-index files from the
packfile odb backend. This also makes git_pack_file objects open their
backing failes lazily in more scenarios, since the multi-pack-index can
avoid having to open them in some cases (yay!).

This change also refreshes the documentation found in src/odb_pack.c to
match the updated code.

Part of: #5399

Introduce GIT_ASSERT macros