fix: work around the Triton buffer-store miscompile per kernel

[WangLingxun]

Why

The AMD Triton backend can emit a buffer_store_dwordx4 whose data VGPRs are redefined by a later instruction with no intervening s_waitcnt vmcnt, so the store writes whatever the clobbering instruction left in those registers. Nothing faults and nothing warns; a few percent of the kernel's output elements are simply garbage. Inductor's SwiGLU backward fusion trips this and poisons the MLP weight gradients, which diverges TorchTitan training. Megatron runs the same inductor-generated kernels, but there the corruption tends to stay inside the normal numeric range, so a finite loss curve does not prove a clean run.

What changed

New patch: primus/core/patches/triton_bufops_war_patches.py. Wraps triton.compile, scans each kernel's emitted AMDGCN for the hazard, and recompiles only the affected kernels with buffer ops disabled. Keying on the machine code means there is no kernel or architecture allowlist to maintain. Registered for both torch backends at the build_args phase, gated only on being a ROCm GPU; non-HIP kernels are skipped per compile. It carries no off switch of its own, since the recompile is a correctness fix and a kernel that does not need one is left alone anyway.

AMDGCN_USE_BUFFER_OPS=0 is deliberately not set globally: the hazardous kernels lose nothing without buffer addressing, but the Primus-Turbo grouped-GEMM kernels spill without it and carry no hazard, so the global switch costs ~18% on MoE recipes for kernels that never needed fixing. Per-kernel selection measured 299.9 vs 299.0 TFLOP/s per GPU on deepseek_v2_lite-BF16, against 244.6 with the global switch.

The patch also appends to TORCH_COMPILE_CACHE_KEY_TAG, since on an FX graph cache hit inductor never calls triton.compile — without invalidation a cache filled before this lands would keep serving hazardous binaries. This is the only environment variable the PR touches, and it is PyTorch's own.

Divergence detection. tests/utils.py now requires every logged loss and grad norm to be finite, because a numerically diverged run still exits 0 and still prints the completion marker. tools/auto_benchmark/metrics.py marks a run with non-finite metrics invalid instead of publishing its throughput.

Tests. 18 tests for the hazard criterion and the ROCm gate, 14 for the divergence parser on both backends' log formats. All static text analysis: no GPU, no Triton.

Docs. A workarounds section in performance-tuning.md and one row in environment-variables.md.

Test plan

• 32 unit tests, ~0.2 s
• MI300X: hazardous kernel fixed and clean kernel untouched, residual 0
• MI355X: E2E green on both backends; the pattern also appears there, though an isolated numeric A/B of the matching kernel came back bit-identical
• CI E2E on both backends