`.lines()` strips line endings, so callers tracking byte offsets
need to re-add the `\r\n` or `\n` length manually.
Extract the repeated inline pattern into a reusable helper.

* Parse `diff --git` extended headers
* split multi-file git diffs at `diff --git` boundaries

Compat test for also `git apply`.

Unlike unidiff,
gitdiff produces patches for empty file creations/deletions
(`0\t0` in numstat)
because they carry `diff --git` + extended headers even without hunks.

Binary files (`-\t-\t`) are skipped in gitdiff mode for now.

* Added types representing both literal and delta Git binary patches
* Added a parser for the `GIT binary patch` format.

This doesn't include the patch application
(which will be added in later commits)

The implementation is based on

* Specification from <https://diffx.org/spec/binary-diffs.html>
* Behavior observation of Git CLI

- Add `Binary::Keep` variant (now the default) to `ParseOptions`
- Add `PatchKind::Binary` variant for binary patches
- Parse `GIT binary patch` payload via `parse_binary_patch`
- Handle `Binary files ... differ` as `BinaryPatch::Marker`
- Add `extract_file_op_binary` for file ops without ---/+++ headers

The API was stabilized in 1.73.
The lint was added in 1.93.

This is required for a MSRV bump to 1.75

This is a preparation for binary diff application support.

* Git binary patch is compressed by zlib hence flate2
* zlib-rs (which is the most performant zlib backend)
  requires MSRF 1.75.0+ hence the bump.

* Add base85 encoder/decoder and Git delta format decoder.
* Wire them into `BinaryPatch::apply() and `apply_reverse()`
  for decoding zlib-compressed, base85-encoded binary payload.

These are feature-gated behind the `binary` feature.

Now both tests require `binary` Cargo feature.

Preparation for `PatchSet::parse_bytes(&[u8])` support.

No behavior change.

Assumption: Header lines are always ASCII
(with `core.quotePath=true`, git's default).

Move dispatch logic to free functions,
so we can have private trait bound on them.

* `next_gitdiff_patch`
* `next_unidiff_patch`
* `Iterator::next` -> `next_patch`

We have the assumption that file path from hunk hader is UTF-8
(This is actually not true but we can leave it for future fix)

Only the `str` constructor and Iterator impl are exposed for now.
`parse_bytes` support comes in a follow-up commit.

Avoid lossy UTF-8 conversion of diff output
so non-UTF8 content round-trips correctly.

No more skips in GitDiff mode (except submodule of course)