[FIX] Fix dimensionality of probabilities from values

[PrimozGodec]

Issue

SGD (same than some other models) outputs wrong shapes of probabilities when those retrieved from values (when skl_learner do not provide probabilities). The issue is caused by the one_hot encoding function that does not consider a number of values in the domain.

SGD (skl_learner) for some of the losses provides probabilities that are not used by Orange since it uses custom predict function.

Description of changes

• one_hot function changed to accept dim parameter which tells what must be a resulting number of columns
• __call__ function for model calls one_hot with a dim parameter that has a number of values in the domain as a shape.
• SGD model now uses a predict function from SklModel (which correctly retrieve probabilities from skl_learner) if possible.

Includes

• Code changes
• Tests
• Documentation

[codecov]

Codecov Report

Merging #4629 into master will decrease coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #4629      +/-   ##
==========================================
- Coverage   83.65%   83.63%   -0.02%     
==========================================
  Files         281      276       -5     
  Lines       56269    55457     -812     
==========================================
- Hits        47072    46384     -688     
+ Misses       9197     9073     -124     

[lanzagar]

There could be one more test for one_hot which would check the case when dim is too low (since we check for that case in the code).

Also the use of one_hot in Orange/widgets/data/utils/histogram.py could be updated to use dim (removing the need for a lengthy comment and the custom fix).