fix(diff): read raw diff records field by field - #2292

`Diff._handle_diff_line` parses `git diff --raw -z` output by splitting
it at every NUL followed by a colon, and `Diff._index_from_raw_format`
called it once per line of stdout. With `-z`, git writes paths verbatim,
so a path may start with `:` or contain a line feed, and either one
breaks the record framing:

- A path starting with `:` is taken as the start of a new record. Its
  remainder is parsed as metadata, the resulting `ValueError` is raised
  inside the stdout pump thread, and the caller gets back a `DiffIndex`
  with a pathless entry followed by nothing: every later entry is
  dropped, and only a log message records the failure. For a commit
  adding `:planted`, `docs/b` and `setup.py`, `commit.diff(NULL_TREE)`
  returned a single entry with `a_path == None`.
- A path containing a line feed is truncated at the line feed. If the
  rest of the path contains a colon, it is parsed as a separate record:
  `docs/a\n:100644 100644 111... 222... M` produced an entry `docs/a`
  plus a fabricated `M` entry whose blob ids came from the file name.

Paths are chosen by whoever authored the commit. A tool that uses
`Commit.diff` to learn which files a commit touches can therefore be
made to miss changes, or to see changes that were never made.

`_index_from_raw_format` now collects all of stdout before parsing, as
`_index_from_patch_format` already does, and `_handle_diff_line` reads
the NUL-separated fields in order: a metadata field, then one path, or
two for copies and renames. A single trailing line feed is still
stripped so the existing non-`-z` fixtures parse as before.

The new test in `test/test_diff.py` fails before the change and passes
after it. `ruff check`, `ruff format --check` and `mypy` are clean. In a
full clone the suite has 6 failures (`test_diff_with_staged_file`,
`test_index_mutation`, `test_empty_repo`, `test_rebasing` and the two
`test_actor_get_uid_laziness_*` tests), and all 6 also fail without this
change.