Find a solution to avoid CUDA OOM during diarization

[chainyo]

By implementing the SpeakerClustering model directly in the project, I'm looking to solve CUDA OOM problems during diarization.

This PR:

• Split diarization sub-modules into multiple files under wordcab_transcribe/services/diarization/*.
• Re-implement the code proposed in the NVIDIA NeMo package for Multiscale Auto-tuning Spectral Clustering with only the first segmentation and clustering steps (not an entire MSDD pipeline, beneficial?).
• Makes the diarization process more modular and comprehensive.
• Adds a dynamic multiscale strategy for long audio files to avoid CUDA OOM.
• Make us close to avoid installing the nemo package (which adds a lot of constraints on Python and PyTorch versions...).