Support CUDA wheels?

[leofang]

Currently, CUDA Python loads NVRTC using standard system paths, for example on Linux (cf man ld.so) we do:

cuda-python/cuda/_cuda/cnvrtc.pyx.in

Line 56 in 2be0aac

handle = dlfcn.dlopen('libnvrtc.so.12', dlfcn.RTLD_NOW)

However, it means unless users purposely add site-packages to LD_LIBRARY_PATH, which is a frown-upon practice, the nvidia-cuda-nvrtc-cuXX wheel (ex: cu12) installed in the user environment will not be used. This issue is for tracking the status of this enhancement.