Improve robust diarization fusion

[desh2608]

Summary

• add raw co-active duration as an opt-in greedy label-mapping affinity
• add equal system weighting, median local speaker cardinality, and recording-level support for tied label votes
• document the reference-free CPU recipe and its AMI development/holdout results
• add focused tests for intersection mapping, median cardinality, and global tie-breaking

Why

The existing DOVER-Lap example scores 19.86% DER on the bundled AMI mix-headset RTTMs. Normalized IoU mapping, rank-derived system weights, rounded-mean cardinality, and inclusive tie handling can overvalue short coincidences or one-system overlap outliers.

The new options remain training-free, reference-free during fusion, and CPU-only. All original defaults remain unchanged.

Results

Split	Default DER	Robust DER
EN2002 + IS1009 development	24.19%	22.87%
ES2004 + TS3003 held out	14.92%	14.56%
All 17 bundled recordings	19.86%	18.99%

Full-set robust components: 10.27% miss, 1.72% false alarm, and 6.99% confusion.

Validation

• python -m pytest -q: 3 passed
• default AMI output remains byte-identical to egs/ami/rttm_dl_test (SHA-256 cd4cdac8428d714e6ebcad96adb1238e4dfc6a61701fe8b31edf2cef0d2d269a)
• robust AMI output scores 18.99% DER with spyder
• focused Black, Flake8, and git diff --check checks pass