More granular top features selection

[brsnw250]

Description

Right now, feature selection transforms selecting the top-k best features as is. But even among these top-k entries, there may be garbage ones, for example, with a low relevance score. 
Maybe we can add additional handles to control the selection process.
Also, it is worth considering a case where the user wants to just filter features by scores. So it might be helpful to provide these handles to the user as well.

Thoughts

We can provide a handle that will apply additional selected feature filtration in the BaseFeatureSelectionTransform._transform method.
This handle depends on the particular selection method implementation and may be optional to implement. For example, it might be additional filtration by some threshold value.

Related

• Generation of prepared transforms after feature selection #481