Move to Dask Distributed

[mauch]

Make the pipeline work with Dask Distributed.

This mostly attempts to avoid saving dataframes to Pandas were possible and also avoids loading dataframes from Pandas. Exceptions to this and their reasons are commented with NOTEs.

Also - where possible avoid persist or pausing the computation with wait until the end of each pipeline step (ie. A wait is issued only at the end of association, new_sources, forced_extract and after source statistics in finalise). Sometimes persist is necessary - and the reasons for this is also commented with a NOTE.

Ready to go now I think.

Still to do:

• Forced Extract
• Finalise
• Tests

[ddobie]

Just some minor stuff, but looking good otherwise.

[ddobie]

Looks good overall. Just some nitpicky stuff about docstrings throughout. mkdocs automatically generates doc pages for each function based on the docstrings, so it's good to keep them as complete as possible.

Once you tidy that up it's good to go!