A multi-pack index uses raw oid data, use a byte array to index
into them.

The index contains entries with raw oid data, use a byte array for the
raw entry data.

A packfile contains arrays of raw oid data, use a byte array to index
into them.

The commit graph contains arrays of raw oid data, use a byte array to
index into them.

Now that oids are type-aware, they use their type to understand how many
bytes to copy.  Some callers may need to copy the raw bytes of the
object id.

This is equivalent to a memcpy that is a little more semantic.

We will talk about "raw" oids as untyped blobs of data; use a name for
the comparison function that is in keeping with that.

Create an object id from raw data instead of casting.

Don't write the object id structure, write its raw oid data.

We explicitly want to write on the id data, not the beginning of the
object data, which may contain other information in the future.

We explicitly want to hash on the id data, not the beginning of the
object data, which may contain other information in the future.

Don't assume that a `git_oid` is a particular size; allocate
`sizeof(git_oid)` instead.

A tree entry previously pointed directly into the object id within the
tree object itself; this is useful to avoid any unnecessary memory copy
(and an unnecessary use of 40 bytes per tree entry) but difficult if we
change the underlying `git_oid` object to not simply be a raw object id
but have additional structure.

This commit moves the `git_oid` directly into the tree entry; this
simplifies the tree entry creation from user data.  We now copy the
`git_oid` into place when parsing.

midx: Fix an undefined behavior (left-shift signed overflow)

Fix crash when regenerating a patch with unquoted spaces in filename

tests: support flaky stat

The 32-bit ARM QEMU builds are flaky when running `lstat`. Disable those
testing `lstat`'s `st_size` temporarily.

diff: don't stat empty file on arm32 (flaky test)

sha256: support dynamically loaded openssl

There was a missing check to ensure that the `off64_t` (which is a
signed value) didn't overflow when parsing it from the midx file. This
shouldn't have huge repercusions since the parsed value is immediately
validated afterwards, but then again, there is no such thing as "benign"
undefined behavior.

This change makes all the bitwise arithmetic happen with unsigned types
and is only casted to `off64_t` until the very end.

Thanks to Taotao Gu for finding and reporting this!

fetch: support OID refspec without dst

Our CI test infrastructure virtualizes arm32 in docker, which is a
sometimes imperfect situation. In `diff::workdir::can_diff_empty_file`,
avoid the stat to ensure that the file is zero bytes; there is an odd
issue running in qemu when emulating arm32 that we should skip.

SHA256: add a SHA256 implementation backend

gcc (mingw) warns when you cast the result of `GetProcAddress`; cast the
results to `void *` before casting them to the actual result.

When GIT_SHA1_WIN32 or GIT_SHA256_WIN32 is used, ensure that we test
both CryptoNG ("cng") and CryptoAPI.

Adding SHA256 support prompted an overdue refactoring of some of the
unnecessary complexity around the CNG/CryptoAPI abstraction.

Add support for a SHA256 hash algorithm, and add the "builtin" SHA256
hash engine (from RFC 6234).

The more generic GIT_ERROR_SHA allows for SHA256 errors as well as SHA1.

Eliminate the `sha1` folder, move it down into `hash` so that future
impelmentations can share common code.

Remove the "generic" implementation; it should never be used; it only
existed for a no-dependencies configuration, and our bundled sha1dc
satisfies that requirement _and_ is correct.

Benchmarking suite

Support the ability to create a refspec that is a single object ID
without a destination.

clone: update bitbucket tests