Audit and update build/test requirements for cuda.parallel and cuda.cooperative

[shwina]

The dependencies listed for cuda.parallel are incomplete:

• lit is required by libcudacxx, which in turn is configured when we first invoke cmake.
• numpy is an unlisted dependency. Currently, I think it's picked up transitively through numba - but should be listed explicitly.

1. At minimum, lit and numpy should be added as dependencies.
2. I think we should move where we list these dependencies from setup.py to pyproject.toml. lit in particular should be available before the package is installed and thus would need to be listed here.
3. We should similarly audit the dependencies of cuda.cooperative.

FYI, @rwgk as I know some of this stuff is probably in flux as part of #3201.