Use a patched version of `CUDA_Driver_jll` to avoid requiring setting environment variables

[apozharski]

This would probably involve backporting something like https://github.com/casadi/polyglot/blob/main/scripts/patch-cuda-driver-jll.py

or maintaining our own .patch file and using patch at build time.