[RadeonFlow] flydsl mxfp4 a4w4 MoE: bring gemm1/gemm2 to parity with the HIP backend

[ColorsWind]

Summary

Brings the flydsl-ported mxfp4 (a4w4) MoE gemm1/gemm2 kernels on gfx950 up to the
HIP backend's performance: closes the large-M (prefill) gap to parity while keeping
the decode/mid-batch lead. Numerically identical (cos ≥ 0.999991 vs HIP, =1 at large M).

Changes

ABI (both kernels + their wrappers)

• Raw-pointer kernel ABI: pass each tensor's data_ptr() (int64) instead of a
  full memref/fx.Tensor descriptor. Kernels assume contiguity and derive sizes from
  compile-time constants, so descriptor shape/stride fields are dead weight; bare
  addresses let the kernarg s_loads coalesce. ~7% faster at decode, parity by M≈64.

gemm2 (aiter/ops/flydsl/kernels/mxfp4_gemm2.py)

• Backend-managed K-loop waitcnt (BM128 nonatomic): drop the hand-tuned
  inline-asm s_waitcnt vmcnt(N) and let the LLVM waitcnt pass own it. The inline
  asm was opaque to that pass, which double-inserted its own waitcnt — the large-M
  scheduling bubble.
• Epilog output-address strength reduction: hoist the one runtime row*N_OUT
  i64 multiply out of the per-element store loop; per-element offsets become
  compile-time constants folded into the store address.
• Unsigned index division for the non-negative grid/tile/count index math
  (drops signed-division sign-correction SALU).

gemm1 (aiter/ops/flydsl/kernels/mxfp4_gemm1.py)

• Hardware exp2 in silu_mul (rocdl.exp2 → v_exp_f32) instead of the
  software math.exp2 expansion (matches HIP's __expf).

Measurement

Per-gemm device time via a cudagraph microbench: capture N back-to-back gemm
launches into one hipgraph, replay, and divide the event-timed elapsed by N
(isolates the gemm from the surrounding pipeline). Exclusive GPU, several
interleaved trials, reported as the in-run flydsl/hip ratio.

Results

Kimi-K2.5 TP=4 shape (NE=385, D_HIDDEN=7168, D_INTER=512, TOPK=9), gfx950,
per-gemm device time, flydsl/hip ratio (lower = faster than HIP):

M	gemm2 before	gemm2 after	gemm1 before	gemm1 after
4	0.92	0.94	0.93	0.93
256	0.93	0.93	1.07	1.01
2048	0.94	0.93	1.03	1.00
4096	1.11	1.00	1.07	1.05
16384	1.10	1.03	1.07	1.04

• gemm2: faster/parity through M=2048; large-M deficit ~10% -> ~0–3%.
• gemm1: faster/parity through M=2048; large-M deficit ~7% -> ~4%.

Notes

• Numerically faithful to the [PERF] MXFP4 (a4w4) MoE backend for gfx950 #3470 HIP source (same e8m0 Even rounding); cos
  preserved at all M.
• Scope is the two kernel files only.

TODO

• Continue closing the remaining large-M (prefill, M >= 4096) gap to HIP: gemm1
  ~4% (MFMA/K-loop bound, K = D_HIDDEN = 7168 dominates), gemm2 ~3% at M = 16384.