perf: optimize KMeans centroid recomputation with thread-local parallel accumulators

[hushengquan]

What

Optimizes KMeansAlgoFloat::to_kmeans centroid recomputation step.

Closes #6369

Changes

Replaced the old parallel-scan approach (each of P cores scans all N data points, filtering by centroid range) with a thread-local parallel accumulation pattern:

1. par_chunks splits data into P chunks (one per rayon thread)
2. Each thread accumulates into a private centroid buffer — zero write contention
3. reduce merges all thread-local buffers into the final centroids
4. Centroid normalization remains parallel over k clusters

Complexity comparison

	Old	New
Data reads	O(N × P)	O(N)
Merge overhead	—	O(k × dim × P)
Write contention	None (disjoint slices)	None (private buffers)

Benchmarks (release, Apple M-series)

Config	Old	New	Speedup
N=131K, k=256, dim=128	3.53 ms	0.48 ms	7.4x
N=524K, k=1024, dim=256	9.98 ms	3.88 ms	2.6x
N=2M, k=4096, dim=256	45.95 ms	24.87 ms	1.9x

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!