[MXFP4] Add scale generation utils

[dsikka]

Summary:

• Adds utilities required to convert scales to be mxfp4 compliant
• MxFp4 scales have the following properties:

1. Scales are rounded to Powers of 2
2. Converted to exponents
3. Stored in uint8

Reference: https://github.com/vllm-project/vllm/blob/main/tests/quantization/reference_mxfp4.py

Added functions:

• Add utilities to generate powers of 2 through round_to_power_2
• This is called by generate_mxfp4_scales which will be called by calculate_qparams. generate_mxfp4_scales will also convert the scales to exponents and store as uint8
• Add convert_mxfp4_exp_scale to convert the exponent scales back to dense such that they can be applied to the weight / activations during QDQ

Testing

• Added unit and some e2e testing
• Will potentially update depending on the vLLM integration, specifically the requirement to -2 when generating the exponent

[brian-dellabetta]

Some nice bit shift wizardry 🧙 , tests look well-covered though

A few nits

[brian-dellabetta]

🔥 🔥 🔥

[rahul-tuli]

Couple very small nits, but LGTM!