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

Initial commit: Speaker Diarization System