Use proper coordinates in the pymc models, store whitened data, compute WAIC and LOO in fit object.

[farr]

Yep. I'm going to merge this into master right now (Harrison wants it), but leave the MR open so that @maxisi can review it.

[farr]

Leaving this alone for the moment. It seems that the string coordinates for the IFO don't interact well with NetCDF compression; see arviz-devs/arviz#2079 . Two options:

• Give up the good IFO and mode names (sigh).
• Turn off compression (sigh).
  What do you think @maxisi @HarrisonS-Phys ?

[farr]

OK. Everything works now. The standard I have settled on is that coordinates labeling the IFO are all UTF-8 encoded byte strings (i.e. b'H1'); and when modes are specified in the mchi series of models, their coordinates are UTF-8 encoded byte strings with the format b'{p}{l}{m}{n}'. This (byte strings rather than UTF-8 Python strings) is required because UTF-8 strings of fixed length do not take up a fixed number of bytes (due to some unicode code points requiring multiple bytes to encode), and thus the coordinate objects can't be compressed when converting to NetCDF.

I'm going to merge into master once the tests pass, but leave the pull request open so you all can look at it.

[farr]

(The failure was just a timeout, so I'm going to merge.)