feat: support diarized Corti transcript input with speaker-aware analysis

[Dan Engel (dwe2501)]

Summary

Adds end-to-end support for diarized Corti transcripts: import, speaker-aware display, per-speaker evaluation, diarization error detection, and a diarization accuracy metric.

What changed

Phase 1 — Speaker-aware transcript import & display

• New tympany/diarize.py parser auto-detects Corti WebSocket streams and REST /transcripts JSON formats, extracts speakerId/channel, and builds per-word speaker labels
• "Import Corti transcript" input mode on the Create page (paste or upload JSON)
• Speaker labels propagated through Token → DiffGroup → Sample → error rows on the results page
• Speaker column, speaker filter, and speaker-tagged CSV export on results page
• BeWER report view shows speaker labels per alignment example

Phase 2 — Per-speaker evaluation

• run_bewer_diarized() in bewer_eval.py splits the transcript by speaker and runs bewer separately per speaker
• "Evaluate per speaker" checkbox on the Create page (visible with Corti transcript input)
• Per-speaker metrics table (WER, CER, updated WER/CER, ref word count) on results and BeWER report pages
• Per-speaker re-run support (recomputes per-speaker metrics after excluding errors)
• Alignment examples split by speaker in the BeWER report view via _split_ops_by_speaker()

Phase 3 — Diarization error analysis

New tympany/diarize_errors.py with align_segments() detecting 5 error types:

• Speaker mismatch — same time range, different speaker (medium risk)
• Merged turns — multiple ref speakers → one gen segment (medium risk)
• Split turns — one ref segment → multiple gen speakers (medium risk)
• Missing turn — ref segment with no gen counterpart (high risk)
• Extra turn — gen segment with no ref counterpart (high risk)
  New diarization_error classification in categorize.py with its own color, dropdown option, breakdown column, and filter on the results page
  diarization_accuracy() metric (matched turns / total turns, coverage ≥ 50% threshold) displayed on results and BeWER report pages

README — updated with new sections for diarized transcript support, per-speaker evaluation, diarization error analysis, updated error classification table, project structure, and development notes.

Test coverage

• 150 tests pass (12 new test functions across test_diarize.py, test_diarize_errors.py, test_bewer.py, and test_report_render.py)
• Test fixtures include clean diarized transcripts and transcripts with deliberate diarization errors
• All new fields default to empty/None/() for backward compatibility with existing non-diarized analyses

[linear-code]

SR-3009