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 only two files.

Summary

• never select flashinfer, sgl_kernel, or OpenAI triton_kernels binaries on ROCm;
• avoid querying the CUDA driver version on an AMD runtime;
• let existing Triton/PyTorch fallback paths handle those operations;
• add unit coverage proving ROCm bypasses all CUDA-only probes while CUDA keeps the existing probe behavior.

Incremental review scope

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

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

Diff size: 46 additions.

Dependency

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

Validation

• ROCm/CUDA backend-selection unit tests: 2 passed
• gfx1201 Triton attention, norm, and activation fallback tests: PASS
• final combined follow-up test set: 54 passed

Compatibility

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