Add a "drop_similar" argument to the TableVectorizer

[GaelVaroquaux]

It would be great to make it really easy drop the redundant columns in the TableVectorizer (Using the DropSimilar transformer inside the TableVectorizer), by a simple additional argument.

This would have improvements both is speed/memory and maybe in statistical performance.

I realize that it makes the TableVectorizer even more a swiss-army knife than it currently is, but honestly, it's sooo useful and we use it everywhere, even as an element in complex pipelines.