turbo-persistence: faster key block sorting (#100335)
## Summary
`KeyOnly` key blocks with 4-byte and 8-byte keys were stably sorted as
`&PendingEntry`, reading both keys through their entries on every
comparison. Each key is now read once into a dense `(key << bits) |
index` integer, and those integers are sorted unstably. The index keeps
equal keys in their original order, preserving the stable order needed
for tombstones. In the original profile, this roughly halved the sort of
a full 4-byte-key block (6.96 → 3.78 µs).
This also adds tests for `MergeIter`, covering ordering, newest-first
equal keys, multi-value groups, input errors, and randomized comparisons
against a sorted reference.
AMQF fingerprint construction is unchanged. I compared `sort_by_key`
against the existing `sort_unstable_by_key` there; on a full compaction
of 16Mi 8-byte-key / 4-byte-value entries over 32 commits, the
difference wasn't statistically significant (5.66 s vs 5.90 s, `p =
0.22`).
## Verification
- `cargo test --features mmap` in `turbopack/crates/turbo-persistence`
- `cargo fmt -- --check`
- `cargo clippy --features mmap --all-targets -- -D warnings`
<!-- NEXT_JS_LLM -->
<!-- fleet af34fc25-1888-4720-a67e-66c65755551a -->
---------
Co-authored-by: vercel-fleet-prod[bot] <318278635+vercel-fleet-prod[bot]@users.noreply.github.com>
Co-authored-by: Luke Sandberg <210140+lukesandberg@users.noreply.github.com>
