Evaluate dask-cuml for 0.8

[JohnZed]

[cjnolet]

Opening this back up because pending CI issues are dragging on and I don't want to lose sight of this for the release.

So far I have verified locally w/ the libcumlmg that test_linear_regression.py and test_nearest_neighbors.py succeed when run.

Unfortunately, the following two issues arose in CI:

• The random projection algorithm is doing a binomial sampling w/o replacement on host, which is pulling in a log gamma function that was built with a newer version of glibc than the one that comes w/ centos (centos has version 19, this was built w/ 23). This likely results from the use of a different libstdc++ lib when linking than what is used at runtime. We are still verifying but believe this is the main issue.
• The most recent build of dask-cuda has an error which is causing the knn test to fail. This is upstream from us and, according to Matthew Rocklin, is upstream from dask-cuda as well.

Once we have confirmed it, I will make an issue #1 so that it can be fixed properly before the 0.8 release.

[cjnolet]

Aside from upstream problems in the dask-cuda project, this is good to go