[Kernel] Add optimized 4-wave MXFP8 GEMM kernel for gfx950

[aris134]

Motivation

Add an optimized 4-wave MXFP8 GEMM kernel for gfx950 targeting training workloads, where operand B is not pre-shuffled. The implementation follows the HipKittens 4-wave execution schedule while adapting it to FlyDSL, including a dedicated scale pipeline with host-prepacked MFMA-ready scale operands, explicit hot-loop schedulers, and ROCDL scheduling controls to preserve the intended memory/compute overlap.

Technical Details

• Add a new 4-wave MXFP8 GEMM kernel for gfx950.
• Implement a two-page ping-pong LDS pipeline for A/B operands.
• Implement a separate scale pipeline using host-prepacked MFMA-ready scale operands loaded directly into VGPRs.
• Add dedicated hot-loop schedulers to guide instruction ordering through the steady-state pipeline.
• Pin accumulators to fixed AGPRs to eliminate unnecessary AGPR↔VGPR transfers and reduce spills.
• Use explicit ROCDL scheduling controls (sched_vmem, sched_dsrd, sched_mfma, sched_barrier, s_waitcnt) to preserve the intended instruction schedule.

Test Plan

• Run the full MXFP8 shape sweep by invoking the test harness with --sweep.
• Each shape performs a standalone correctness check before its timed benchmark.
• Benchmark performance on MI355X (gfx950) using the same sweep.

Test Result

• Correctness: 68/68 benchmark shapes passed.
• Performance (MI355X / gfx950):
  • Mean: 2209.22 TFLOPS
  • Median: 2223.65 TFLOPS
  • Max: 2554.10 TFLOPS

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.