mixed-mla: replace AITER reference with pure pytorch baseline

[theodorechapman]

This replaces the mixed-mla reference implementation with a pure PyTorch baseline so the task no longer depends on AITER compilation, which has been slowing down submission time for the AMD competition

• kept the same task-facing structure:
  • generate_input(...)
  • ref_kernel(...)
  • check_implementation
• preserved the fp8 Q / fp8 KV decode path in pure PyTorch
• kept kv_data generation for bf16, fp8, and mxfp4

This submission is not bitwise identical to the old mla_decode_fwd path.

However, When tested on a MI355X, the code passed comprehensive tests against the current AITER reference with max error in the 3.8e-06 to 1.5e-05 range. Got the same when I submitted it through popcorn as a custom kernel.

Only caveat is that this implementation is much slower then the AITER reference. AITER reference sits around 200µs whereas this Pytorch implementation sits around 4000

[danielhua23]

keep these introduction

[theodorechapman]

Had this fixed locally haha, just pushed

[danielhua23]

initial ok for me, but we need check if the two impls are exactly the same, since this is weekedn, we will answer you in two days