fix(build): honor explicit and versioned CUDA toolkits

[HaileyStorm]

Summary

• resolve an explicit absolute CUDA_HOME before importing torch cpp_extension
• otherwise select the exact versioned /usr/local/cuda-X.Y toolkit matching PyTorch, then an exact matching nvcc on PATH
• keep complete prebuilt-cache startup tolerant of an unrelated PATH compiler while retaining strict checks at every JIT/build entry
• apply the same exact-toolchain gate to source builds and kernel-cache wheel compilation

Scope

Related to #140, but deliberately narrower: this does not change FreeToken dependency pins, the supported PyTorch/driver/GPU matrix, or the default CUDA 13 requirements. It is not a blanket CUDA 12 or V100 support claim.

I checked the active dependency/toolchain PRs before opening this. #26 handles uv dependency groups, while #253 changes the Volta/cu128 dependency line and cache architectures; neither implements this runtime/build-time toolkit resolver.

Verification

Linux x86_64 host with PyTorch 2.11.0+cu130:

• PYTHONPATH=python python -m pytest -q -p no:cacheprovider tests/kernels/test_toolchain.py tests/kernels/test_kernel_cache_version.py — 25 passed
• python -m py_compile on the five changed Python modules/tests — passed
• git diff --check — passed
• live compiler discovery resolved PyTorch CUDA 13.0 to /usr/local/cuda-13.0
• explicit /usr/local/cuda-12.8 against cu130 failed closed with the expected exact-release mismatch

No GPU execution, model load, or performance claim is involved.