[TheNetherWatcher]

Summary

Timeline.co_iter re-scans every segment of other that starts before the current segment ends, so it is O(n·m). It sits under crop, get_overlap, Annotation.__mul__ and therefore under every pyannote.metrics computation. Profiling DiarizationErrorRate on a 4k-segment file: 92% of the time is spent in co_iter, with ~70M Segment.intersects calls. Timeline.segmentation had the same pattern via support.overlapping(middle).

This PR replaces both with single-pass sweeps:

• co_iter: keep a sorted list of candidate segments of other; stop scanning at the first one sorted after Segment(end, end) (same bound as before); drop candidates that start before the current segment and do not intersect it, since they cannot intersect any later segment either.
• segmentation: track the running maximum end time of support segments sorted before Segment(middle, middle) and starting at or before middle, since middles only increase.

Output is unchanged: same pairs, same order.

Benchmark

Synthetic diarization (4 speakers, ~15% overlapped speech), pyannote.metrics 4.2.0, Python 3.11, Apple M-series:

segments	DER before	DER after	IER before	IER after
1,000	0.30s	0.06s	0.26s	0.04s
4,000	3.94s	0.28s	3.35s	0.18s
8,000	14.8s	0.58s	13.1s	0.39s

DER with collar=0.5, skip_overlap=True at 8k segments: 27.0s → 0.74s. All metric values are identical to the last digit. 32k segments now takes ~2.5s for DER.

Correctness

• New tests compare both methods against naive reference implementations on random timelines, including ties, nested, touching and sub-SEGMENT_PRECISION segments (these are needed to exercise the precision-dependent branches of Segment.intersects).
• Differential fuzzing against the previous implementation: 78k random cases, including Segment.set_precision(...), zero-length and invalid segments, all identical.
• Lean 4 proof that both sweeps return exactly what the original loops return, for any total order on times and any rounding of end - SEGMENT_PRECISION: https://github.com/TheNetherWatcher/lean-perf-proofs/tree/main/pyannote-core-sweep

Related: #14 (also about the cost of support/crop on large annotations).