Add MXFP casting kernels from triton Repro

[drisspg]

Summary

They have recently published alot of good upcast and downcast kernels: https://github.com/triton-lang/triton/blob/main/python/triton_kernels/triton_kernels/numerics_details/mxfp.py
We should update the ones we have in AO and bench against Inductor

[DehuaTang]

Hi, @drisspg :

I found that the results of triton's implementation of mxfp4 and torchao's implementation of mxfp4 are not strictly comparable. May I ask why this happens?

[drisspg]

Hmm do you have any scripts or code pointers for this?

[DehuaTang]

Hi, @drisspg :

I found that torchao uses the ties round to the nearest even digit strategy (https://github.com/pytorch/ao/blob/main/torchao/prototype/custom_fp_utils.py#L11)when fp32 to fp4. But in triton, it uses the round up strategy (https://github.com/triton-lang/triton/blob/main/python/triton_kernels/triton_kernels/numerics_details/mxfp.py#L700).

Do you know why triton took this approach?

Thanks！