[sdba] Optimizations and fixes

[aulemahal]

Pull Request Checklist:

• This PR addresses an already opened issue (for bug fixes / features)
  • This PR fixes Usage question: How to use multiple members as samples in sdba? #810
• Tests for the changes have been added (for bug fixes / features)
• Documentation has been added / updated (for bug fixes / features)
• HISTORY.rst has been updated (with summary of main changes)
• bumpversion (major / minor / patch) has been called on this branch
• Tags have been pushed (git push --tags)
• The relevant author information has been added to .zenodo.json

What kind of change does this PR introduce?

• Optimizes map_blocks when using coords backed by cftime.datetime objects. Dask needs some sort of hash for each input and that tokenization is quite slow for cftime np.ndarrays because it goes through pickle.dumps (see Performance : speeding up pickling of cftime arrays Unidata/cftime#253). Also, CFTimeIndex are slow to create because of a check that cycles through all elements of the array. With many chunks and large time dimensions, this takes a significant portion of the call time (excluding the actual computation). Here we add an option to encode cftime coordinate into integers (+ attributes), like in a netCDF, saving time on the cftime handling. The decoding is done within the task itself (thus in parallel).
• Optimizes adapt_freq to skip computation on all-nan blocks.
• Fixes nbu.vecquantile when rnk is NaN.
• Merges all step of DQM.adjust in a single map-block-wrapped function. Once again to reduce the number of tasks in dask's tree. However, the adjust tasks as large and will trigger warnings about garbage collection taking a lot of cpu time. I haven't found any optimization for this part yet.
• Better dtype conservation. Fixes at many places, including in map_blocks to set the template variables to the largest dtype of the inputs.
• Modify map_blocks and map_groups to consider dimensions in Grouper.add_dims. By default, functions wrapped with map_groups will assume Grouper.DIM AND Grouper.ADD_DIMS are all reduced, unless main_only=True is passed.
• "constant" Extrapolation of adjustment factors in qm_adjust is now using values just over and under the max and min of the target data, instead of ±np.inf. There was a bug I had while adjusting precipitation in ESPO-R that was solved by this. I'm not completly sure why...

Does this PR introduce a breaking change?

I think not, but this PR was so iterative over a long time, I'm not 100% sure.

Other information:

All those optimizations and fixes were done while processing "scénarios génériques v2". Except the add_dims thing.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB