Avoid using a stale mmap() by tracking the filestamp.

Since mmap() on windows locks the file, and this map is long-lived,
just copy the file into memory instead. This enables us to keep the
lookup() paths the same, while avoiding the downsides of mmap() on
windows.

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Co-authored-by: Edward Thomson <ethomson@github.com>

Currently ref lookups require loading the entire packed-refs file into
a hashmap in memory. For repos with large numbers of refs this can be
painfully slow.

This patch replaces the existing lookup code and instead mmap()'s the
packed-refs file and performs a binary search to locate the ref entry.
Git uses a similiar approach.

The old hash table codepath is still used for unsorted packed-refs files.

This patch also fixes a minor bug where the "peeled" trait is never
parsed correctly from the packed-refs header.

Introduce `git_object_rawcontent_is_valid`

Users may want to validate raw object content; provide them a function
to do so.

Return `GIT_EINVALID` on parse errors so that direct callers of parse
functions can determine when there was a failure to parse the object.

The object parser functions will swallow this error code to prevent it
from propagating down the chain to end-users.  (`git_merge` should not
return `GIT_EINVALID` when a commit it tries to look up is not valid,
this would be too vague to be useful.)

The only public function that this affects is
`git_signature_from_buffer`, which is now documented as returning
`GIT_EINVALID` when appropriate.

Minor code cleanups

Fix repo init when template dir is non-existent

Instead of using the project-specific `libgit2_SOURCE_DIR` and
`libgit2_BINARY_DIR` variables, use `CMAKE_SOURCE_DIR` and
`CMAKE_BINARY_DIR`.

The filebuf functions should use hashes directly, not indirectly
using the oid functions.

The tests that examine sha1 behavior (including collision detection)
should test against the hash functionality directly, not indirectly
using the oid functions.

Remove some unnecessary includes from utility code.

util: minor cleanup and refactoring to the date class

Instead of `git__date`, just use `git_date`.

cmake: re-enable WinHTTP

Ensure that we `enable_testing()` at the top-level CMakeLists.txt or
else we'll need to navigate within the build directory to the correct
place in the hierarchy to run `ctest`.  Now we can `ctest` at the
top-level again.

tag: set validity to 0 by default

`git_tag_name_is_valid` needs to set validity to 0 when it
short-circuits.