[bghira]

SM86 Ampere GPUs were rejected by SimpleTuner's attention controller and native entry point, while Ada retained the older wide-head layouts. Sync KohakuFA-CuTe c0786dd and enable the existing CuTe backend for SM86 in both strict and automatic training modes.

• Port Ada D256/D512 tile selection, paired QK/dP score producers, packed DS/P stores, and wider SM86/SM89 D512 operand copies. Include the accompanying SM120 cache-policy and causal D128 query updates.
• Extend native compatibility checks, automatic dispatch, centered autograd handling, and capability error messages. Test SM86 acceptance and SM80/SM87 rejection, plus training/evaluation policy transitions.
• Generalize warm-cache and changed-input CUDA graph regressions across SM86/SM89/SM120. Expand compiled projection-layout and gradient-cancellation coverage, and test optimized selection against masked/unaligned layouts.
• Update the vendored revision notice, WebUI attention tooltip, and existing option documentation in English, Chinese, Japanese, Portuguese, Spanish, and Hindi. No new backend option is required.

Validation:

• .venv/bin/python -m unittest discover -v -f tests/: 7,091 tests run, 323 skipped, no failures.
• RTX 3090 (SM86), RTX 4090 (SM89), and RTX 5090 (SM120): 31 targeted checks passed per device, with no skips, in isolated attention-component environments using PyTorch 2.11.0+cu130, CuTe DSL 4.8.1, TVM FFI 0.1.15.post0, and Diffusers 0.40.0.
• Hardware coverage includes FP16/BF16 outputs and Q/K/V gradients, saturation and tiny-gradient preservation, masks/GQA/strides, eager and compiled backward, cache reuse, changed-input CUDA graph replay, key centering, strict/automatic training dispatch, and SimpleTuner/Diffusers integration.
• Device VRAM returned to the pre-test baseline after each test process exited. All validation rentals were released.
• Vendored kernels.py matches the pinned upstream file byte-for-byte. Native C++ matches after namespace adaptation, and imported regression methods match after import-path integration. Pre-commit formatting, import sorting, lint, and whitespace checks passed.

These are correctness and dispatch checks; this PR does not claim newly measured attention latency or full training throughput.