fix(rocm): gate CUDA-only optional backends on ROCm

[zihaomu]

Draft follow-up to #132. This branch is based on the #132 head. Until #132 merges, GitHub Files changed also includes the foundation diff; the incremental review scope is three files.

Summary

• never select flashinfer, sgl_kernel, OpenAI triton_kernels, Marlin, or other NVIDIA donor binaries on ROCm;
• explicitly route NVFP4 auto/triton requests to Triton on ROCm;
• fail clearly for forced marlin/flashinfer requests without probing or importing donor packages;
• avoid querying CUDA driver/capability paths on an AMD runtime;
• remove the unused driver_hip_version() helper;
• retain existing Triton/PyTorch fallback paths and CUDA selection behavior.

Incremental review scope

Relative to zihaomu:feat/rocm-rdna3-rdna4-foundation:

• python/freetoken/kernel/backend.py
• python/freetoken/moe/nvfp4_backends.py
• tests/kernels/test_backend.py

Diff size: 107 additions, 11 deletions.

Dependency

• Depends on feat(rocm): add RDNA3 and RDNA4 runtime foundation #132.
• Independent of the JIT and RCCL follow-ups.

Validation

• backend-selection tests: 8 passed
• ROCm NVFP4 tests cover auto, triton, marlin, and flashinfer;
• tests assert ROCm never executes package probes or donor imports;
• gfx1201 Triton attention, norm, and activation fallback tests: PASS.

Compatibility

The CUDA path continues to probe and select the same optional packages. This remains Draft until #132 establishes the shared ROCm foundation.

[zihaomu]

Dependency and architecture refresh (2026-09-16): rebased this PR onto the conflict-resolved #132 head c2af4c6 and ported the CUDA-only backend gates to the current quantization architecture. The old python/freetoken/moe/nvfp4_backends.py remains deleted; ROCm now gates flashinfer, sgl_kernel, and the current vLLM probe, while tests exercise the new layers.quantization kernel selector for Triton fallback and forced Marlin/b12x rejection. Validation: tests/kernels/test_backend.py (5 passed), adjacent MiniMax NVFP4 selector test (1 passed), AST and git diff --check. New head: e35a0ec; GitHub reports clean/mergeable. No GPU hardware tests were rerun for this conflict refresh.