Migrate as much as possible to pyproject.toml

[vyasr]

Description

I migrated as much build logic as possible to pyproject.toml for cudf, dask_cudf, and custreamz. A couple of notes:

• I didn't move cudf_kafka because it is currently still using a pure setuptools build with Cython files. We may as well do the pyproject.toml migration there after we have also done the scikit-build conversion since there's a lot of logic that can't be removed from setup.py until then. I'm also not entirely confident that everything will work as expected if we move data to pyproject.toml without switching to scikit-build based on issues like the ones mentioned in this SO post about Cython and project data stored in pyproject.toml.
• I could get rid of setup.py entirely for custreamz at this point, but I expect that we'll run into a host of additional issues when we do eventually move to requiring pip install X instead of python setup.py X across RAPIDS and I'd rather deal with all of those at the same time with our main packages rather than using an esoteric package like custreamz as the test case.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[galipremsagar]

Should we have the python 3.9 classifier consistent across all .toml's? Either we drop it everywhere or keep it incase python 3.9 packages are making a comeback.

[vyasr]

Good question. We removed it in rmm so I'm going to continue removing it everywhere else during this transition for now. We can readjust later if we want to.

[galipremsagar]

LGTM, some non-blocking comments.

[galipremsagar]

Veirifed the backend dispatching works fine 👍

[vyasr]

/merge