Fixes for CVE 2018-11235

Added note about Windows junction points to the differences from git document

We might modify caches due to us trying to load the configuration to figure out
what kinds of filesystem protections we should have.

We still compare case-insensitively to protect more thoroughly as we don't know
what specifics we'll see on the system and it's the behaviour from git.

When dealing with `core.proectNTFS` and `core.protectHFS` we do check
against `.gitmodules` but we still have a failing test as the non-filesystem
codepath does not check for it.

Any part of the library which asks the question can pass in the mode to have it
checked against `.gitmodules` being a symlink.

This is particularly relevant for adding entries to the index from the worktree
and for checking out files.

This is so we have it available for the path validity checking. In a later
commit we will start rejecting `.gitmodules` files as symlinks.

We may take in names from the middle of a string so we want the caller to let us
know how long the path component is that we should be checking.

We want to reject these as they cause compatibility issues and can lead to git
writing to files outside of the repository.

These will be used by the checkout code to detect them for the particular
filesystem they're on.

These can't go into the public API yet as we don't want to introduce API or ABI
changes in a security release.

Given a path component it knows what to pass to the filesystem-specific
functions so we're protected even from trees which try to use the 8.3 naming
rules to get around us matching on the filename exactly.

The logic and test strings come from the equivalent git change.

It checks against the 8.3 shortname variants, including the one which includes
the checksum as part of its name.

This lets us check for other kinds of reserved files.

Otherwise we would also admit `..\..\foo\bar` as a valid path and fail to
protect Windows users.

Ideally we would check for both separators without the need for the copied
string, but this'll get us over the RCE.

If the we decide that the "name" of the submodule (i.e. its path inside
`.git/modules/`) is trying to escape that directory or otherwise trick us, we
ignore the configuration for that submodule.

This leaves us with a half-configured submodule when looking it up by path, but
it's the same result as if the configuration really were missing.

The name check is potentially more strict than it needs to be, but it lets us
re-use the check we're doing for the checkout. The function that encapsulates
this logic is ready to be exported but we don't want to do that in a security
release so it remains internal for now.

cmake: resolve libraries found by pkg-config 

refdb_fs: enhance performance of globbing

global: adjust init count under lock

Fix GCC 8.1 warnings

With the recent change of always resolving pkg-config libraries to their
full path, we do not have to manage the LIBGIT2_LIBDIRS variable
anymore. The only other remaining user of LIBGIT2_LIBDIRS is winhttp,
which is a CMake-style library target and can thus be resolved by CMake
automatically.

Remove the variable to simplify our build system a bit.

Libraries found by CMake modules are usually handled with their full
path. This makes linking against those libraries a lot more robust when
it comes to libraries in non-standard locations, as otherwise we might
mix up libraries from different locations when link directories are
given.

One excemption are libraries found by PKG_CHECK_MODULES. Instead of
returning libraries with their complete path, it will return the
variable names as well as a set of link directories. In case where
multiple sets of the same library are installed in different locations,
this can lead the compiler to link against the wrong libraries in the
end, when link directories of other dependencies are added.

To fix this shortcoming, we need to manually resolve library paths
returned by CMake against their respective library directories. This is
an easy task to do with `FIND_LIBRARY`.

Worktrees can be made from bare repositories

We were previously conflating any error into GIT_ENOTFOUND, which might
or might not be correct. This fixes the code so a config error is
bubbled up, as well as preserving the semantics in the face of
worktree-repositories

docs: add documentation to state differences from the git cli

Sanitize the hunk header to ensure it contains UTF-8 valid data

The diff driver truncates the hunk header text to 80 bytes, which can truncate
4-byte Unicode characters and introduce garbage characters in the diff
output. This change sanitizes the hunk header before it is displayed.

This mirrors the test in git: https://github.com/git/git/blob/master/t/t4025-hunk-header.sh

Closes https://github.com/libgit2/rugged/issues/716

examples: ls-files: add ls-files to list paths in the index

Since GCC 8.1, the compiler performs some bounds checking when
copying static data into arrays with a known size. In one test,
we print a format string of "%s/sub%02d" into a buffer of 64
bytes. The input buffer for the first "%s" is bounded to at most
63 characters, plus four bytes for the static string "/sub" plus
two more bytes for "%02d". Thus, our target buffer needs to be at
least 70 bytes in size, including the NUL byte. There seems to be
a bug in the analysis, though, because GCC will not account for
the limiting "%02" prefix, treating it as requiring the same
count of bytes as a "%d".

Thus, we end up at 79 bytes that are required to fix the
warning. To make it look nicer and less special, we just round
the buffer size up to 80 bytes.

Since version 8.1, GCC will do some automatic bounds checking
when printing static content into a buffer with known size. The
bounds checking doesn't yet work quite right in all scenarios and
may thus lead to false positives. Fix one of these false
positives in refs::normalize by simplifying the code.

Our provided callback function `threadid_cb(CRYPTO_THREADID
*threadid)` sets up a unique thread ID by asking pthread for the
current thread ID.  Since openssl version 1.1,
`CRYPTO_THREADID_set_numeric` is simply a no-op macro, leaving
the `threadid` argument unused after the preprocessor has
processed the macro. GCC does not account for that situation and
will thus complain about `threadid` being unused.

Silence this warning by using `GIT_UNUSED(threadid)`.

Our global initialization functions `git_libgit2_init()` and
`git_libgit2_shutdown()` both adjust a global init counter to determine
whether we are the first respectively last user of libgit2. On
Unix-systems do not do so under lock, though, which opens the
possibility of a race between these two functions:

    Thread 1                            Thread 2
                            git__n_inits = 0;
    git_libgit2_init();
    git_atomic_inc(&git__n_inits);
    /* git__n_inits == 1 */
                                        git_libgit2_shutdown();
                                        if (git_atomic_dec(&git__n_inits) != 0)
                                            /* git__n_inits == 0, no early exit here */
                                        pthread_mutex_lock(&_init_mutex);
                                        shutdown_common();
                                        pthread_mutex_unlock(&_init_mutex);
    pthread_mutex_lock(&_init_mutex);
    init_once();
    pthread_mutex_unlock(&_init_mutex);

So we can end up in a situation where we try to shutdown shared data
structures before they have been initialized.

Fix the race by always locking `_init_mutex` before incrementing or
decrementing `git__n_inits`.