Convert to using xarray and dask.

[ReadingClouds]

I have converted to use xarray; this simplifies much of the code and also facilitates the use of dask to parallelize the processing. Of note:

• get_data enhanced to provide true grids, at least for primary data (so we now have x_p, x_u, y_v as well as z and zn as coords). Q for reviewers: do we still need '_on_w' etc. appended to variable name. This makes access quicker but loses some of the xarray benefits.
• dask is sensitive to chunking - I have only tested on my laptop. Beware small chunks - I managed to slow doen processing of a small file by a factor of 10! Note for reviewers: experiment with sf.subfilter_setup['chunk_size':] - currently 2**22 numbers.
• dask can also get its knickers in a twist saving one field after another - I added a 3 second pause after each which seems to solve the issue. Note for reviewers: experiment with sf.subfilter_setup['write_sleeptime].
• I have saved deformation as a 5D array (time, x, y, z, i, j) - this gets very big (25 GB with George's data) and netCDF access doesn't like the size. xarray/dask works fine with output. However, an alternative is my original solution, to save 9 arrays and keep in memory as a dictionary. I have partially implemented this. Q for reviewers: which do you prefer?
• Alanna - see new version of your main program.

[trjr]

I have been unsuccessful with this in multiple configurations on JASMIN. It always seems to fail on the d.compute(...) with TypeError: slice indices must be integers or None or have an __index__ method. It's not clear to me where the problem is or how to find out, and I have sent a message to JASMIN support asking whether any special considerations are required on their system.

Has anyone had any luck with this code? Any suggestions?