chore:optimize next_coalesced() by tracking removed ids in a HashSet

[caicancai]

This PR speeds up State::next_coalesced() by avoiding an O(k) linear scan for every candidate when checking whether a request was already marked for removal. Previously, scanning a window of n candidates with k already-marked ids could cost O(n·k) (worst-case O(n²)). We keep the Vec<(offset, id)> for removals, but add a HashSet for O(1) average membership checks, bringing the scan closer to O(n). Coalescing behavior is unchanged.

[codspeed-hq]

Merging this PR will not alter performance

✅ 1138 untouched benchmarks
⏩ 1265 skipped benchmarks¹

---

_{Comparing caicancai:perf/io-coalesce-hashset (91257b9) with develop (cfd9116)}

Footnotes

1. 1265 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[caicancai]

@joseph-isaacs Sorry, the issue of cargo clippy failing has been fixed.