docs: fix CUDA Attention skill-doc guidance to match bottom-right causal + fully-masked errata

[titaiwangms]

What

Docs-only correction to the hand-written agent skill doc .agents/skills/cuda-attention-kernel-patterns/SKILL.md. No code or kernel change.

Why

The skill doc gave guidance opposite to the behavior shipped in #28958 (which implements the onnx/onnx#8068 errata). It would actively misdirect agents working on the ONNX-domain Attention op:

• is_causal=1 + external/static KV cache (nonpad_kv_seqlen, no past_key): the doc called this spec-invalid and told models to use is_causal=0. It is in fact valid and supported, using bottom-right (offset-aware) alignment — offset[b] = nonpad_kv_seqlen[b] - q_sequence_length (clamped >= 0). Corrected, including the bottom-right key-range bullet that was written backwards.
• Fully-masked query rows → Y = 0 on BOTH EPs (CPU Bug-2 guard + CUDA ZeroFullyMaskedRowsKernel). Removed the stale mean(V) 'spec reference' claim and the now-resolved cross-EP TODO.
• qk_matmul_output mode-3: documented fully-masked row = 0 (mandated, consistent with Y), and that CUDA returns NOT_IMPLEMENTED for mode-3 (it is a CPU-only path).

Grounding

Matches the landed kernel: core/providers/cuda/llm/attention.cc (bottom-right dispatch, mode-3 NOT_IMPLEMENTED), attention_mask_impl.cu (ZeroFullyMaskedRowsKernel), core/providers/cpu/llm/attention.cc (Bug-2 row-zeroing). See #28958 and onnx/onnx#8068.

Risk

Documentation only — no build/runtime impact.

[titaiwangms]

Folding this docs change into #28958 since it documents the same kernel behavior (bottom-right is_causal, fully-masked Y=0, mode-3=0) — keeping it atomic with the kernel change for review/merge. Closing in favor of #28958, which now carries the SKILL.md correction as a dedicated docs commit (1964d9b).