[Feature] Add speaker diarization for multi-speaker videos

[Edmon02]

Summary

Currently the pipeline assumes a single speaker. Add speaker diarization to handle multi-speaker videos with different voice clones per speaker.

Tasks

• Integrate pyannote.audio or similar diarization model
• Map each segment to a speaker ID
• Support multiple reference audio files (one per speaker)
• Clone each speaker's voice independently
• Track speaker transitions in the pipeline metadata

Considerations

• GPU memory management with multiple voice models loaded
• Speaker embedding similarity threshold for clustering
• Graceful fallback to single-speaker mode