[cuda backend] skip fully-masked KV blocks calculation in SDPA

[Gasoonjia]

This PR skip attention calculations entirely masked by

1. sliding-window
2. causal via start_n > max_seq_pos (lower tranigle of attention calculation)

result:

prompt len	before	after	gain (vs before)	llama.cpp (cuBLAS prefill t/s)	gain (vs llama.cpp)
256	729.9	1065.7	+46%	1150.5	−7.4%
512	935.2	1582.7	+69%	1609.9	−1.7%
1024	1116.9	2024.5	+81%	1601.2	+26.4%
2048	1238.0	2325.5	+88%	1573.7	+47.8%

Prefill +46-88% all lengths; decode remain the same;
Surpass llama.cpp by 48% on 2k input while on par (-7.4%) on short prompt
Profling on gemma4-31b: SDPA takes 58.1% ->18.5% of e2e prefilling time.
Numerically bf16-exact vs dense+mask (unit test).

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20198

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures, 2 Unrelated Failures

As of commit ac2968b with merge base a79f3e4 ():

NEW FAILURES - The following jobs have failed:

• pull / test-mcu-cortex-m-backend / linux-job (gh)
  RuntimeError: Command docker exec -t 8b58e5b92eb460bc08d1a174571ba0c7c435cb9768e69702c8d6cce2ca6ff178 /exec failed with exit code 1
• Test CUDA Builds / unittest-cuda / linux-job (gh)
  backends/cuda/tests/test_int4_dispatch.py::TestMultiLayer::test_two_layer_mlp
• Test CUDA Windows Export and E2E / test-model-cuda-windows-e2e (facebook, dinov2-small-imagenet1k-1-layer, non-quantized) / windows-job (gh)
  Process completed with exit code 1.

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• Test CUDA Builds / export-model-cuda-artifact (SocialLocalMobile, gemma-4-31B-it-HQQ-INT4, quantized-int4-tile-packed) / linux-job (gh) (trunk failure)
• Test CUDA Builds / export-model-cuda-artifact (SocialLocalMobile, Qwen3.5-35B-A3B-HQQ-INT4, quantized-int4-tile-packed) / linux-job (gh) (trunk failure)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[linux-foundation-easycla]

• ❌ The email address for the commit (ac2968b) is not linked to the GitHub account, preventing the EasyCLA check. Consult this Help Article and GitHub Help to resolve. (To view the commit's email address, add .patch at the end of this PR page's URL.) For further assistance with EasyCLA, please visit our EasyCLA portal and chat with our support bot.