Improve SuperVectorizer

[LilianBoulard]

Implements various improvements in the SuperVectorizer, such as :

• Missing values are not systematically imputed anymore (Fixes Missing values replaced by the SuperVectorizer #178)
• Type casting and per-column imputation are now learnt during fitting, and applied during transform

[LilianBoulard]

Updated original comment to describe the changes of d5f6193
I still have a bit of work to do, especially improving the tests and documenting the new properties.

[GaelVaroquaux]

Partial review. I'll finish this ASAP

[GaelVaroquaux]

OK, I finished going through all the code, and I put in all my comments. This PR will be good to go.

[GaelVaroquaux]

For standard keywords, yes, but not for type hinted ones.

Thank you for correcting me. My knowledge is getting too old :)

[alexis-cvetkov]

I finished reviewing the PR:

• we need to update CHANGES.rst
• maybe add a test to ensure that fit_transform(X) is equivalent to fit(X).transform(X)
• is there a reason to define a separate encoder for string column and categorical columns in the parameters of the SuperVectorizer ?

[GaelVaroquaux]

LGTM

[GaelVaroquaux]

The questions of Alexis are very relevant. Once they are addressed, I am +1 for merge.

[LilianBoulard]

Alexis' questions are addressed by the last commits, thanks for the feedback !

[alexis-cvetkov]

Merging !