[FEA] Allow libraries that implement __array_ufunc__ to override CUDAUFuncDispatcher

[gmarkall]

Details / notes:

From Allow libraries that implement __array_ufunc__ to override CUDAUFuncDispatcher on the Numba Discourse.

There was a PR implementing this for the CPU target: https://github.com/numba/numba/pull/8995

A related issue on the Awkward issue tracker: https://github.com/scikit-hep/awkward/issues/3179

This is to support using Coffea on CUDA.

cc @ianna @lgray

Initial thoughts:

I am not able to dig deep into this right now, but there are a couple of additional potential hurdles for the CUDA target:

• The CUDA Ufunc mechanism doesn't yet support dynamic Ufuncs, which seem to be a part of the implementation of the PR for the CPU target.
• The __array_ufunc__ mechanism seems CPU-centric - will we need to define a new __cuda_array_ufunc__ mechanism for this to be practical? c.f. __cuda_array_interface__ vs. __array_interface__.

[gmarkall]

See #37, which is a better-written request.