fix: reduce diarization speaker label noise

[loookashow]

Summary

This PR reduces speaker-label noise in two places:

• refines Spectral Clustering labels with spherical centroid reassignment
• expands silhouette refinement around the BIC estimate instead of only trying k, k+1, k+2
• converts overlapping embedding windows into a non-overlapping final timeline with local majority smoothing

Benchmark

Evaluated VoxConverse locally:

Before:

• Weighted DER: ~8.46%
• Mean DER: ~8.45%
• Median DER: ~2.72%
• Exact speaker-count match: 111/216

After:

• Weighted DER: 5.16%
• Mean DER: 6.65%
• Median DER: 2.37%
• Exact speaker-count match: 117/216

Notes

This improves the benchmark substantially, but it does not fully solve noisy speaker switching on real meeting audio. There are still cases where speaker labels flip too often inside a single speaker turn.

Validation

• python -m ruff check src/diarize/clustering.py src/diarize/__init__.py tests/test_diarize.py
• python -m pytest tests/test_diarize.py -q

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!