Feat/flydsl attention authoring levers

[amd-wangfan]

Summary

• Add FlyDSL guidance for fused multi-GEMM attention kernels (authoring_attention_levers.md), plus the gfx942 MLA backward SOTA card and a draft flydsl_fused_attention_backward expert skill (training only, not auto-applied).
• Qualify the gfx942/gfx950 “fewer waves / 4-wave interleave” prior for operand-bound attention backward; leave the GEMM ranking unchanged.
• Record transferable mechanisms (LDS layout, AGPR shuttling vs scratch, XCD dispatch order) and a scoped exception in benchmarking methodology for isolated kernel variant sweeps.

Scope: perf_knowledge/ only. Same 54 operators, 226 backend cards. No inference-path or workflow behavior change.

Test plan

• Static-validate the new skill: python perf_knowledge/expert_skills/_contribute/validate_skill.py flydsl_fused_attention_backward --static.
• Regenerated SOTA registry/matrix match the committed files (226 cards; restore CRLF after _gen_registry.py).
• New/amended cards have ## Sources. L0 is unaffected (docs-only). No L1: skill stays draft and is not auto-applied.

Benchmark results

MLA 192/128 causal varlen backward, gfx942 (times in µs):

	config	fwd	bwd	fwd+bwd	total
A	dv=128 unpadded: ASM fwd + CK bwd	192.7	1717.9	1910.7	1809.8
B	V padded to 192: CK fwd + ASM bwd	430.0	1091.9	1521.9	1535.8
C	dv=128 unpadded: ASM fwd + FlyDSL bwd	192.4	649.1	841.4	891.1

• C vs A: fwd+bwd 2.27×, backward-only 2.65×, whole-step 2.03×
• C vs B: fwd+bwd 1.81×, backward-only 1.68×, whole-step 1.72×

Note: The generated kernels have already been included in the aiter PR, ROCm/aiter#4875