fix: detect CUDA by system install, not torch or driver cap

Worker is a separate subprocess that loads quantfunc.dll via ctypes and
does not import torch, so torch.version.cuda does not describe what the
DLL will actually load. Likewise, nvidia-smi only exposes the driver's
max-supported CUDA, not the installed toolkit.

Reorder detect_cuda_major() to: CUDA_PATH → installed toolkit on disk
→ nvidia-smi driver cap → torch (last-resort guess). Add Linux coverage
for /opt/cuda and /usr/lib/cuda, and resolve unversioned symlinks like
/usr/local/cuda → cuda-12.4 via os.path.realpath.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>