[FEA] Use nogil wherever possible when making calls to libcuml

[jcrist]

Currently long running cuml calls that dispatch into libcuml don't release the GIL. For example, a call to UMAP.fit will block any other python threads, preventing the python interpreter from doing other work while it waits for the libcuml call to return.

A use case I had recently was starting a logger for system resources as a background thread to run while UMAP.fit was running. Since the fit call doesn't release the GIL, the background thread never gets to run until the fit call returns, defeating the purpose.

If there's a known reason why we don't release the GIL in these calls, we should document that somewhere. And if there isn't, then I'd argue that we should try and release it in more places.

[jcrist]

Looking through the codebase, it looks like we do release the GIL for some but not all estimators. I'd propose that we should probably be more uniform here, and if any shouldn't release the GIL we should add a comment next to the cdef as to why not.

[cjnolet]

I was about to respond with "I thought we were using nogil everywhere" but then I looked at a few of the estimators and realized it was never added to those. I definitely think we should create a feature request out of this issue and strive to use nogil everywhere possible.