Allow linking to CUDA dynamically instead of dlopen()ing it at runtime