Add SDPA edge case test coverage

[hwanseoc]

@coderabbitai ignore

Summary

• Add deterministic packed SDPA zero-sequence coverage for both seq_len_* and prefix-sum cu_seq_len_* inputs.
• Cover zero Q, zero KV, and zero QKV sequences; cu_seq_len_* cases are forward-only because backward plumbing is not supported.
• Cover bottom-right causal empty rows and top-left causal empty columns without bias injection.
• Exercise varied all-negative attention scores in both FP16 and BF16 with nonzero backward gradients.
• Reuse the canonical SDPA execution/reference harness through narrow input and result hooks.

Validation

• A100 SM80, backend 9.30: 7 passed, 3 skipped (packed seq_len_* backward unsupported)
• H100 SM90, backend 9.30: 10 passed
• B200 SM100, backend 9.30: 10 passed
• Existing SDPA harness regression: test_sdpa_random_fwd_L0[test1] passed

[jhjpark]

Is it possible to combine the edge case tester with test_mhas_v2.py? I think it would be nice if test_mhas_v2 could cover it by itself.

[hwanseoc]

Is it possible to combine the edge case tester with test_mhas_v2.py? I think it would be nice if test_mhas_v2 could cover it by itself.

I'd prefer keeping the deterministic edge cases separate.

They already use the same tester and reference implementation as test_mhas_v2

Integrating them into random generation would either weaken coverage or add constrained combinations and unnecessary test expansion.

[Anerudhan]

@hwanseoc will submit after necessary waives. LGTM

[Anerudhan]

@cudnn-ci-bot run

[cudnn-ci-bot]

🚀 Running mirror pipeline

Branch: cudnn-gh/pr-328-be17996
Pipeline: 56518344

[hwanseoc]

@cudnn-ci-bot run

[cudnn-ci-bot]

🚀 Running mirror pipeline

Branch: cudnn-gh/pr-328-be17996
Pipeline: 56673003

[hwanseoc]

@cudnn-ci-bot run

[cudnn-ci-bot]

🚀 Running mirror pipeline

Branch: cudnn-gh/pr-328-be17996
Pipeline: 57024864