Patches may have no hunks when there's no modifications (for example,
in a rename).  Handle them.

Test that we can create a diff file, then parse the results and
that the two are identical in-memory.

Like `git_patch_to_buf`, provide a simple helper method that can
print an entire diff directory to a `git_buf`.

Parse diff files into a `git_diff` structure.

Test with some postimages that actually grow/shrink from the
original, adding new lines or removing them.  (Also do so without
context to ensure that we can add/remove from a non-zero part of
the line vector.)

Parse values up to and including `\377` (`0xff`) when unquoting.
Print octal values as an unsigned char when quoting, lest `printf`
think we're talking about negatives.

`oid_strlen` has meant one more than the length of the string.
This is mighty confusing.  Make it mean only the string length!
Whomsoever needs to allocate a buffer to hold a string can null
terminate it like normal.

Handle prefixes (in terms of number of path components) for patch
parsing.

When a text file is added or deleted, use the file names from the
`diff --git` header instead of the `---` or `+++` lines.  This is
for compatibility with git.

We may have parsed binary data, set the `SHOW_BINARY` flag which
indicates that we have actually computed a binary diff.

Now that `git_diff_delta` data can be produced by reading patch
file data, which may have an abbreviated oid, allow consumers to
know that the id is abbreviated.

No need to replicate the old_file/new_file members, or plumb them
strangely up.