Right now the compiler libraries are dynamically linked at the build time of the CCCL C library. However they should be "run-time linked," i.e. dlopen'd, for a few reasons:

1. It makes it impossible to support Python packaging
   • These DSOs can be installed in many ways and the location varies depending on the installation method, which EPIC: Path finder for CUDA components cuda-python#451 will address but it won't work for C libraries
2. It makes it harder to support 2 CUDA major versions ([FEA]: Match cuda.parallel support matrix with CuPy #2554), since the soname in each major version is different
   • We can of course build the C library twice, and publish two wheels cuda-parallel-cu11 & coda-parallel-cu12, but generally this can really be avoided and shipping cuda-parallel is possible by carefully designing the DSO loading mechanism
3. It could cause ImportError at cuda.parallel's import time
   • We could hide the import of the C library somehow (not impossible especially after we fix cuda.parallel: move away from using ctypes #3854).

2 & 3 can be worked around but not 1. Fixing this issue would also align with other Python projects (cuda-bindings, cuda-core, nvmath-python).