Test SelectKBest + StandardScaler pipeline

[Pierre-Bartet]

When SelectKBest is used before a StandardScaler inside a pipeline with $k$ smaller than the number of features, the conversion fails:

model = make_pipeline(SelectKBest(k=3), StandardScaler(), LogisticRegression())

This PR fixes this problem by using the concat shape calculator for selectors and adding a correct one instead.

[Pierre-Bartet]

@xadupre I fixed the issue (just for SelectKBest, not the other selectors), but before I go further I want to be sure I understand properly what was the initial intention of the code I changed.

All selectors such as SelectKBest use the concat shape calculator instead of something closer to the array_feature_extractor one. This look like a mistake to me, so I unregistered the selectors from the concat shape calculator, and use a slight variation of the array_feature_extractor one instead.

What bothers me is that in the end the calculate_sklearn_array_feature_extractor is never called during the test I added (and never was).

So my two questions are:

1. Was the use of the concat shape calculator for the selectors just a mistake, or was there some deeper intention here?
2. Is it intended that with or without my fix, calculate_sklearn_array_feature_extractor is not used when converting SelectKBest into an array_feature_extractor?

[xadupre]

You should remove the register and run tests/test_sklearn_feature_selection_converters.py to see which unit tests require it. shape_calculator is a kind of legacy but it is used to compute the output shape and type of a model without converting it so that the conversion of the next estimator can start before the consersion of the previous is done. In practice, this never happens.

[Pierre-Bartet]

No tests/test_sklearn_feature_selection_converters.py test fail when doing that, but other tests fail elsewhere.
So from what I understand, the answer to question 2. is "Yes".
I guess this is also the case for question 1, so I'll just proceed and do the same for all selectors.

[Pierre-Bartet]

@xadupre done