perf: compute HNSW level counts after build

[BubbleCal]

No description provided.

[github-actions]

Code Review

Summary: This PR optimizes HNSW index building by removing atomic operations from the hot path. Instead of incrementing level counts atomically during each parallel insertion, the counts are computed once after the build completes.

Assessment: LGTM ✅

The change is correct and the optimization is sound:

1. Correctness: num_nodes() is not called during the parallel build phase, so deferring level count computation doesn't affect build behavior. The final compute_level_count() correctly iterates over all nodes.

2. Performance: Eliminates atomic fetch_add operations from the hot insertion path, which should reduce contention in the parallel loop.

3. Testing: The new test test_level_offsets_match_batch_rows properly validates that level offsets are correctly computed and match the final batch size.

No issues found.

[codecov]

Codecov Report

❌ Patch coverage is 97.14286% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
rust/lance-index/src/vector/hnsw/builder.rs	97.14%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[wjones127]

Look good!