Implement MLA decode fwd nh128 a8w8 kernel with FlyDSL.

[ruanjm]

ATT.

[Copilot]

Pull request overview

Adds a FlyDSL implementation of an MLA decode forward kernel specialized for nhead=128 with FP8 Q/KV inputs and BF16 output, plus a Python launcher and a corresponding correctness/perf test driver.

Changes:

• Introduce kn_mla_fwd_decode_m16x8_fp8_fp8 FlyDSL kernel and JIT launcher for the nh=128 FP8/FP8 decode path.
• Add a thin Python dispatcher (flydsl_mla_fwd_decode) that flattens inputs/outputs and launches the specialized kernel.
• Add a new MLA decode test/benchmark script with a PyTorch reference + aiter metadata/reduce integration.

Reviewed changes

Copilot reviewed 2 out of 3 changed files in this pull request and generated 8 comments.

File	Description
tests/kernels/test_mla_decode.py	New MLA decode reference + kernel launch driver (currently not pytest-integrated and has a couple runtime issues).
kernels/mla_fwd_decode_m16x8_fp8_fp8.py	New FlyDSL kernel implementation + JIT launcher for FP8/FP8 nh=128 decode.
kernels/mla_fwd_decode.py	Public Python launcher/dispatcher for the new specialized kernel.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[coderfeli]

Align code style, type, IR usage with other kernels. Try to remove or reduce arith/std_arith and similar native mlir low level uses.

[ruanjm]

Align code style, type, IR usage with other kernels. Try to remove or reduce arith/std_arith and similar native mlir low level uses.

Done