[FEA] Standalone NRT header/library

[brandon-b-miller]

Is your feature request related to a problem? Please describe.
I'd like to be able to compile a c++ portion of my library that leverages the NRT CUDA api to PTX without needing to depend on the python part of numba-cuda and its dependencies. The code could live as a lightweight separate library that could be depended on by numba-cuda itself and any other consumers of the NRT CUDA api.

Describe the solution you'd like
Separate NRT C++ code into its own library, package it and depend on it within numba-cuda

Describe alternatives you've considered
Adding numba-cuda as a build dependency for the package

Additional context
The particular use case are UDFs in cuDF that use memory owning types within the UDF logic, such as strings or columns in a custom groupby aggregation.