Support various Whisper model with Metal backend

[seyeong-han]

Summary

This PR enhances Whisper model flexibility during export while simplifying the runtime interface by removing the model_name argument from the runner.

Changes

Export Scripts Enhancement

• Added model_name argument to export.sh
  • Allows specifying any HuggingFace Whisper model (tiny, base, small, medium, large, large-v3, large-v3-turbo)
  • Defaults to openai/whisper-large-v3-turbo if not specified
• Automatic feature size detection based on model variant
  • Uses 128 mel features for large-v3/large-v3-turbo models
  • Uses 80 mel features for all other models
  • Prevents tensor shape mismatch errors by correctly configuring the preprocessor

Runtime Simplification

• Removed model_name argument from run.sh and main.cpp
  • Hardcoded decoder_start_token_id=50258 for all models
  • Fixes tokenizer compatibility issue where all Whisper models from HuggingFace now use the v3 tokenizer format
  • Eliminates confusion about which model name to pass at runtime

E2E Script Updates

• Updated e2e.sh to support --model-name flag during export
• Simplified run step to no longer pass model name

Documentation

• Comprehensive README updates with model comparison table
• Added examples for different model variants
• Documented mel features and tokenizer configuration

Why These Changes?

1. Export Flexibility: Users can now easily export any Whisper model variant without modifying scripts
2. Correct Preprocessing: Automatic feature size detection ensures the preprocessor matches the model's requirements
3. Tokenizer Fix: All HuggingFace Whisper models now use the updated tokenizer format (token 50257 = <|endoftext|>, token 50258 = <|startoftranscript|>), so hardcoding 50258 works universally
4. Simplified UX: Removing the model_name runtime argument reduces user confusion and potential errors

Testing

Tested with whisper-tiny and whisper-large-v3-turbo models to verify correct transcription output.

cc. @manuelcandales