[FMHA] add flash_attn_interface L2 wrapper and cross-length Q/KV (seqlen_q != seqlen_kv)

[yanguahe]

Motivation

This PR adds two things on top of what is already in main (#683, #685):

1. kernels/flash_attn_interface.py — a new process-level L2 API wrapper
   (flydsl_flash_attn_func) that provides @lru_cache build caching, unified
   shape / dtype / arch validation, automatic cross_seqlen detection, and
   split-K workspace management.

2. Cross-length attention (seqlen_q != seqlen_kv) — the gfx950 DUALWAVE_SWP
   kernel now supports arbitrary Q and KV sequence lengths with a
   bottom-right-aligned causal mask, for both dense and packed-varlen
   (cu_seqlens) inputs.

The accompanying test-harness update merges the four run_config* /
run_splitk_config functions into a single run_attn_config, shares the
reference computation across FlyDSL / aiter_ck / aiter_asm in --compare mode
(eliminating 2 out of 3 expensive reference calls), and exposes all kernel
options as CLI flags.

Technical Details

Changed files (4):

File	Change
kernels/flash_attn_interface.py	new — 371-line L2 wrapper
kernels/flash_attn_gfx950.py	seq_len_kv param + bottom-right causal + cross_seqlen flag
kernels/flash_attn_generic.py	thread cross_seqlen through dispatch; seq_len_kv in fallback guard
tests/kernels/test_flash_attn_fwd.py	unified run_attn_config; shared ref; CLI flags

Test Plan

python tests/kernels/test_flash_attn_fwd.py --iters 100 --compare

on MI355X (gfx950), HIP_VISIBLE_DEVICES=1.

Test Result

Measured on AMD Instinct MI355X (gfx950), bf16+fp16, --compare vs main
across all DEFAULT_CONFIGS (27 shapes x 2 dtype x causal + non-causal = 108
matched configurations), HIP_VISIBLE_DEVICES=1:

• Average +1.1% FlyDSL TFLOPS vs main across all 108 matched configurations.
  Breakdown: bf16/causal +0.7%, bf16/nocausal +1.6%, fp16/causal +1.3%,
  fp16/nocausal +1.0%.
• Peak gain +10.9% at B=2,S=1024,H=64,bf16/nocausal.
• Largest regression -2.9% at B=16,S=8192,H=64,fp16/causal — within
  run-to-run noise for a roofline-pinned config.
• All MaxErr within bf16/fp16 tolerance. black + ruff clean.

Submission Checklist

• black + ruff check clean (line-length 120).
• Accuracy verified on MI355X (gfx950) — all DEFAULT_CONFIGS PASS.
• Performance measured before/after on MI355X (bf16+fp16).
• CI green.