Global NRT, remove dask in favor of stdlib multiprocessing

Additions:

• removes dask in favor of using multiprocessing stdlib. It's not as elegant as the dask code so if you have a preference between them then say something. We don't "need" to make this change

• there's also a comment and try/except here that hint at maybe what went wrong with dask and where it was hiding errors. But the questions are good for us to consider in terms of "graceful handling"