fix: gate CUDA directory checks on GPU vendor to prevent false CUDA detection (#8942)

Container images that install CUDA runtime libraries (e.g., cuda-cudart-12-5
via apt) create /usr/local/cuda-12 directories as a side effect. The previous
code checked for these directories before checking whether a GPU was present,
causing CPU-only hosts to select a CUDA backend that crashes because
libcuda.so.1 is absent.

Reorder checks so CUDA directory existence only refines the capability when
an NVIDIA GPU is actually detected, consistent with the arm64 L4T code path.

Signed-off-by: Sertac Ozercan <sozercan@gmail.com>