[Audio] Support Whisper V3

[kylesayrs]

Purpose

• Support Whisper V3 model

Changes

• Change default whisper model to v3
• Modify preprocessing function to be simpler
• Add dtype conversion to preprocessing function
  • Note that this is only required for feature extractor processors, as they return values which are float types (not just token ids, which work regardless of model dtype)

Follow-ups

• Dtype conversion should theoretically be injected into prebaked dataset pathways as well, although I consider this low priority since we push users towards writing their own data processing functions

Testing

• Quantized Whisper v3 model
• Note that you may have to add ds.cleanup_cache_files() to line 40 in order to overwrite any existing mapping caches

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

[brian-dellabetta]

❤️