Reduce speaker over-segmentation in auto clustering

Add pyannote-first diarization path and tune fallback clustering

Fix stereo duration handling and robust ECAPA model loading

Fix portability, model cache handling, and deploy token safety

Fix LocalStrategy.COPY for speechbrain 1.0.3

Fix from_hparams for speechbrain 1.0.3

Fix cache dir for HuggingFace Spaces

Initial commit: Speaker Diarization System