sdba - Adapt "PrincipalComponents" adjustment

[aulemahal]

I had a chat with @Paattrriicckk about sdba in general, but more specifically about xclim.sdba.PrincipalComponents.

Todo:

• Maybe rename the class? It is not a principal components because all components are used.
• Whether here only or sdba-wide, emphasize that groupings other than "time" and "time.dayofyear + window", can give really strange results. Maybe a warning upon the initialization of the Grouper class?
• Implement the "best orientation" algorithm of Patrick. It is more "brute-force" than the current one (which is from the article), but it ensures the best correlation and allows extraction of some useful diagnostics. Would be an option of the object.
• Evaluate the difference between centring the rotations on the mean (Hnilica) or the median (Grenier). Maybe, allow the user to choose. I am talking of

  xclim/xclim/sdba/adjustment.py

  Line 933 in 1cd5c88

  scenario = ds.ref_mean + ds.trans.dot((ds.sim - ds.vmean), [lbl_M])

  (i.e. : ref_mean vs ref_median)

[Paattrriicckk]

Yes, very good exchanges with @aulemahal. Quick comments:

1. In my own script and related upcomin paper, I now prefer to use "eig", or "eigen" or "eigenvectors" or "rotation", instead of pc (although some code segments still use names like "prcomp").
2. When I cross grouping option "window" with "multi: eig+QM", I get terrible jumps at month-t-o-month interfaces. Tomorrow my M. Sc. student will give me a broader picture on month-to-month jumps across all 18 adjustment alternatives that we test.
3. If two options for determining eigenvectors orientation are coded, tests (e.g., visual inspection of 365-day stretches, with [raw] vs. [eig_mean] vs. [eig_median], or correlation coefficients btw these input and outputs) should reveal what performs best.
4. I'm glad you coded @Balinus 's technique for extremes; this is my next step, but possibly not on the short term due to concurrent priorities.