[bghira]

Sync the vendored CuTe implementation with merged upstream revision 1afb72698c74b3a1b5958d3b80948458c9bb6e84 from KohakuFA-CuTe.

This imports specialized Ada/Hopper causal GQA forward paths, split forward attention for short cross-attention queries, and flattened launch grids when batch × heads exceeds 65,535. Native extension build names now include the C++ source hash, preventing stale binaries when copied checkouts preserve modification times.

Port numerical, saved-statistics, cache/stride, graph replay, launch-boundary, and architecture-specific forward regressions. The comparison script reuses SimpleTuner's existing centered FP64 reference. Update the vendoring notice and all OPTIONS translations.

Validation:

• .venv/bin/python -m unittest discover -v -f tests/ — 7,154 tests completed successfully, with 345 skips.
• Local focused attention tests — 100 tests completed successfully, with 77 skips; native build-cache regression also passed.
• H100 with PyTorch 2.13.0+cu130: 29 attention/forward tests completed successfully, with seven architecture-specific skips; all 11 dispatch configuration checks passed.
• The previous implementation passes 65,535 batch/head groups and fails at 65,536 with CUDA error 9; the imported implementation passes both boundaries, including masked and grouped-query cases.
• Kernel ASTs match upstream after namespace adaptation and two lint-only local variable renames.
• Extreme saturated FP16 logits reproduce a pre-existing H100 precision limit: the baseline and imported implementations have bit-identical outputs/gradients, with approximately 0.96% dQ error against FP64 in the reported case. The split-forward regression compares gradients with the serial CuTe path there; ordinary-input coverage retains FP64 checks.
• All four additional short-query split-forward tests passed on H100, covering FP16/BF16 saved statistics, graph replay, saturated logits against the serial control, and compiled dynamic query-length boundaries.
• All pre-commit hooks passed.

The CPU CI job installs PyTorch-family wheels from the official CPU index and verifies Torch, TorchAudio, and TorchVision imports before test discovery, avoiding mixed CUDA runtime builds in a CPU test environment.