Implement BetterTransformer alternative

[chainyo]

This PR isn't meant to be merged, it illustrates how BetterTransformer could replace faster-whisper for transcription optimization. For being merged, it needs more work on:

• Extra language handling
• Make transcription batch_size adaptative based on the detected hardware
• Using the vocab with the new model implementation
• Handling multi-gpu context
• Rewriting the config and environment variables
• Removing unused API parameters that were specific to faster-whisper
• Update the VadService to avoid using faster-whisper
• Remove faster-whisper from dependencies
• Fix the multi-channel and live implementations
• Update the schemas in the model.py file
• Remove unused imports (a simple lint command should work)
• Fix tests