rerere: Skip mismatched marker sizes like git-rerere does

[rwe]

When parsing conflicts, git-rerere ignores/preserves conflict markers that are not exactly the expected length, which defaults to 7.

For <<<<<<< and >>>>>>> , it also requires them to end in a space, to avoid ambiguity.

See git/rerere.c:is_cmarker for reference + additional commentary.

This updates git-revise's parsing to treat them the same way.

[krobelus]

Looks good.
If the conflict-marker-size gitattribute is set, git merge-file can produce longer conflict markers.
Since git-revise doesn't read gitattributes we'll fail to parse those.
I don't think that's a problem because conflict-marker-size is usually used for files containing >>>>>>> which we previously attempted to parse as conflict. Fixing that is more important.

[rwe]

Yes — I actually have a commit that uses git check-attr to get that value, but I'm hesitant to push it up due to unknown performance impact. If introduced it should probably be batched/cached similarly to how OIDs are handled. If interested, that naïve implementation looks like this (plus some extra notes):

def get_git_attr(repo: Repository, file: Path, attr_name: str) -> bytes:
    # TODO: Batch multiple paths, attrs and cache them. Keep the process open like with cat-file.
    # Result format is:
    #   <path> NUL <attribute> NUL <value> NUL
    result = repo.git("check-attr", "--cached", "-z", attr_name, "--", file)
    # Note that check-attr respects cwd, so if we want to use relative paths,
    # these would be relative too.
    # The response name is exactly the input, not normalized:
    # if we ask for "Foo" vs "foo" vs "./foo", the value we asked for is in the response.
    (resp_path, resp_attr_name, value) = result.split(b"\0", maxsplit=3)[:-1]
    assert os.fsdecode(resp_path) == str(file)
    assert resp_attr_name.decode() == attr_name
    return value


def get_conflict_marker_size(repo: Repository, file: Path) -> int:
    # See ll_merge_marker_size in git/ll-merge.c
    value = get_git_attr(repo, file, "conflict-marker-size")
    try:
        if value != b"unspecified":
            return int(value)
    except ValueError:
        pass
    return DEFAULT_CONFLICT_MARKER_SIZE

[rwe]

Oh, woops, my reply above refers to yet another branch which I did intend to PR, which factors that marker_size value out. I'll push that up.

EDIT: See #101

[mystor]

Thanks for the fix :-)