Add MXFP4 Quantization Support

[zhutao100]

This PR introduces MXFP4 quantization support to the mlx-vlm library, extending the existing quantization capabilities.

Summary of Changes

• New quantization mode: Added --q-mode command-line option with support for both affine (traditional) and mxfp4 (new) quantization modes
• MXFP4-specific constraints: MXFP4 mode enforces group_size=32 and bits=4 parameters to ensure proper operation
• Enhanced quantization function: Updated the quantization workflow to support the new mode parameter and improved quantization predicates
• Testing: Added new test cases to verify MXFP4 quantization functionality
• Robustness improvements: Enhanced tokenizer handling and improved the logic for processing inputs

Technical Details

• Added a new --q-mode argument to the conversion script with choices between affine (default) and mxfp4
• When MXFP4 mode is selected, the code automatically enforces the required parameters (group_size=32, bits=4)
• Modified the quantization utilities to properly support different quantization modes while maintaining backward compatibility
• Improved the get_class_predicate function to handle various quantization scenarios
• Added robustness improvements for tokenizer inputs and audio processing

Testing

• Added tests to verify both affine and mxfp4 quantization modes
• Verified that mxfp4 mode correctly enforces required parameters
• Ensured existing functionality remains intact

This feature allows users to leverage MXFP4 quantization, which can provide benefits in specific use cases, especially for memory-constrained environments while maintaining model performance.

[reneleonhardt]

MXFP4 is awesome, at least on text 😄 @Blaizzy Can you run workflows? They don't run automatically.

[zhutao100]

@Blaizzy This PR is updated to date.

[Blaizzy]

Thanks @zhutao100!

Quick question, since we have mlx-lm as a dependency isn't there a simpler approach of doing this?

For instance, check how mixed quant works.