Fix rare edge case with extremely inbalanced data

[mfeurer]

For dataset 360112 Auto-sklearn would fail because the data would
first be sub-sampled and then contain some classes only once.
In the internal splitting, the StratifiedShuffleSplit would not
be able to split the dataset into train and valid, and would resort
to only a ShuffleSplit. This could put the single sample for a
class into the test set. At predict time we would then miss one class.

This commit creates two new splitters which move a sample from the
test split to the training split if a class does not exist in the
train split.

[mfeurer]

CC @eddiebergman we could use these splitters to simplify subsample_if_too_large, right?

[codecov]

Codecov Report

Merging #1244 (a88d754) into development (63808ef) will decrease coverage by 0.17%.
The diff coverage is 64.94%.

@@               Coverage Diff               @@
##           development    #1244      +/-   ##
===============================================
- Coverage        88.25%   88.08%   -0.18%     
===============================================
  Files              139      140       +1     
  Lines            11034    11128      +94     
===============================================
+ Hits              9738     9802      +64     
- Misses            1296     1326      +30     

Impacted Files	Coverage Δ
autosklearn/evaluation/splitter.py	59.75% <59.75%> (ø)
autosklearn/evaluation/train_evaluator.py	92.39% <93.33%> (-0.03%)	⬇️
autosklearn/util/logging_.py	88.96% <0.00%> (+0.68%)	⬆️
...eline/components/feature_preprocessing/fast_ica.py	97.82% <0.00%> (+6.52%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 63808ef...a88d754. Read the comment docs.

[eddiebergman]

CC @eddiebergman we could use these splitters to simplify subsample_if_too_large, right?

Yes this splitter would probably work in that case too, or at least I see no reason it shouldn't.