compat: provide missing FA3 private forward

[yeahdongcn]

Summary

• add flash_attn_interface._flash_attn_forward only when the provider omits it
• adapt the public flash_attn_func(..., return_softmax_lse=True) result to the
  low-level (output, softmax_lse, S_dmask, rng_state) inference contract
• preserve a provider's native private implementation unchanged and keep the
  patch idempotent

Motivation

vLLM-Omni Ring Attention imports _flash_attn_forward because it needs the
softmax LSE for partial-attention accumulation. The current MATE compatibility
module exposes the equivalent public output+LSE API but not that private
symbol, so the downstream FA3 availability probe evaluates false.

The adapter is deliberately conditional and version-independent. Once MATE or
another provider exposes _flash_attn_forward, torchada leaves the native
object untouched.

Safety boundaries

• require the public callable to explicitly support return_softmax_lse
• translate only the keyword subset used by the Ring FA3 inference path
• fail closed if the public provider does not return at least output and LSE
• return None for dropout-only auxiliary values unavailable from the public
  inference API

Validation

• ruff check src/torchada/_patch.py tests/test_cuda_patching.py
• focused compatibility tests: 2 passed, 212 deselected
• full non-hardware test slice: 410 passed, 17 skipped, 41 deselected
• wheel build: torchada-0.1.81-py3-none-any.whl
• S5000 downstream smoke in
  vllm-omni:minimax-h3-20260815@sha256:23ae27867cd19ce848a27688dba262d0569c67ff3c3b599cc2a429f8ab184a8b:
  • installed this exact commit editable over the image's torchada 0.1.79
  • vLLM-MUSA platform-plugin import captured the shim before vLLM-Omni Ring
    globals; HAS_FA3=True
  • vLLM-Omni's unchanged FA3 selector resolved to
    ring_kernels.fa3_forward
  • BF16 [1,1024,14,128] output and LSE were finite; shim vs the public MATE
    output+LSE API had max absolute difference 0.0
  • real two-rank MCCL Ring Attention completed on both ranks with
    attn_type=fa3, ring_kernel=fa3_forward, and finite
    [1,512,14,128] output