…ormance

`packed::Buffer::binary_search_by` calls `decode::reference` on every
comparison step (~17 per lookup on a 154k-ref store), and each decode
runs `gix_validate::reference::name` -> `gix_validate::tag::name_inner`
plus full hex validation on the hash. The binary search only needs the
name bytes for ordered comparison; the final match in `try_find_full_name`
already re-parses the record through `decode::reference` and surfaces any
parse error.

Introduce `decode::name_at_record_start`, which scans for the
`<hash><space><name><newline>` shape and returns just the name slice
without validating either piece, and use it from the binary search.
Parse-failure detection is preserved via the same flag the previous
implementation set.

On a wide-refs incremental fetch (gitaur AUR mirror, 154k branches)
this is the single largest CPU sink — `gix_validate::tag::name_inner`
shows up at ~14% of total fetch CPU in `samply`, all of it from this
loop. End-to-end on the same fetch the change brings wall time from
~11.0s to ~7.7s (~30%) and user CPU from ~8.1s to ~5.1s (~37%).

Co-authored-by: Claude Opus 4.7 <[email protected]>

- refactor
- [P2] Reject bare CR before accepting a record name — /Users/byron/dev/github.com/GitoxideLabs/gitoxide.skip-validation-in-binary-search/gix-ref/src/store/packed/decode/mod.rs:118-118
  When a packed-refs record contains a bare `\r` that is not followed by `\n` and the lookup is for a different name, this fast path now returns `Some(name)` and leaves `encountered_parse_failure` unset. The full parser rejects the same record because `parse::newline` only accepts LF or CRLF, so this can turn a malformed sorted packed-refs file into `Ok(None)` instead of `Error::Parse`.

Co-authored-by: Sebastian Thiel <[email protected]>