feat(attention): share standalone FA3 varlen and LSE adapter

[yeahdongcn]

Summary

Extract the standalone FlashAttention-3 adapter from the MAGI-2/MUSA reference draft into the shared diffusion attention utilities.

• Support the standalone FA3 package layouts through one cached resolver.
• Normalize output-only and (output, LSE) calls, including return_attn_probs / return_softmax_lse aliases.
• Forward optional native sinks only when the provider explicitly supports them; do not silently emulate unsupported features.
• Reuse this adapter from vllm_flash_attn_varlen_with_lse when the bundled API cannot be imported. Preserve the existing bundled-backend priority and do not redirect explicit FA2/FA4 requests.

No MAGI-2 algorithm, MUSA platform gating, kernel implementation, EP, sampler change or performance configuration is included. No environment-variable inventory changes.

Tests

Changed-file pre-commit: all pass (including Ruff and mypy).

55 CPU contract/attention tests pass in a CPU-only container, including existing version-selection coverage. The two MUSA numerical tests now pass; the two CUDA numerical tests remain not run. Five existing device-dependent attention tests were excluded from this CPU-only run; the initial unrestricted selection failed on accelerator allocation because the container had no GPU.

python -m pytest -q -o addopts='' -p no:cacheprovider \
  tests/diffusion/attention/test_fa3_varlen_adapter.py \
  tests/diffusion/attention/test_flash_attn.py -m cpu \
  -k 'not test_padding_equivalence and not test_fa_vs_sdpa and not test_flash_attn_func_preferred_over_varlen and not test_cross_attn_key_padding_vs_sdpa'

On the MUSA development image, torchada was imported before invoking pytest.main with the above arguments. The MUSA provider validation below covers this adapter only. CUDA runtime validation and model end-to-end/performance remain not run; this PR stays Draft.

MUSA provider validation

Exact tested head: 612bdad0914c3c13214fd2d31d1f391a31eeeacf.

• 2 passed, 24 deselected, 0 skipped, including causal and non-causal packed BF16 GQA with unequal Q/K lengths.
• Both output and packed LSE are checked against independent FP32 Torch references. Existing tolerances are unchanged: output rtol=2e-2, atol=2e-2; LSE rtol=3e-3, atol=3e-3.
• MTT S5000, one 48-SM device, driver 5.2.0-server. Python 3.10.12, torch/torch_musa 2.11.0.post1+musa5.2.0, torchada 0.1.83, MATE 0.2.6, flash_attn_3 0.2.6+musa, vLLM 0.28.0, vllm-musa 0.1.28.
• Resolved provider: flash_attn_3.interface through the installed torchada wrapper; return_attn_probs requests LSE.
• Image: registry.mthreads.com/mcconline/inference/vllm-omni@sha256:7f4f2cd83a88979025ef92968334f6e075eb24d45d02e0c30dbb222f7780d58a.
• Exact checkout installed with SETUPTOOLS_SCM_PRETEND_VERSION=0.28.0 python -m pip install -e . --no-deps --no-build-isolation; import path verified. Install, tests and container exit codes were 0.

After importing torchada, the runner invokes:

pytest.main([
    "-vv", "-s", "-o", "addopts=", "-p", "no:cacheprovider",
    "tests/diffusion/attention/test_fa3_varlen_adapter.py", "-m", "musa",
])

No native-sink, MAGI-2 consumer, full-model, compilation/capture or performance qualification is implied by these two adapter tests. The separate MAGI-2 FA3 consumer still has its own numerical gate; this result does not qualify that branch.

Split boundary

This is the generic adapter part of #7156, which remains unchanged as reference. The later MUSA wiring PR will consume the shared API; it is intentionally not bundled here. BF16 MoE remains in #7206.

[yeahdongcn]

Superseded by #8498, which carries a trimmed version of this adapter (no backend selector, no fa3_fwd_interface candidate).