Right now we make it easy to overwrite _hstack in ColumnTransformer, but I wonder if we should try to be more generic in the first place. I don't think without something like nep 37 we can do fully generic stacking, but maybe we can already allow stacking pandas dataframes?

I can think of few cases where all the transformers output pandas dataframes but you want to output a numpy array in the ColumnTransformer.
This would be a backward incompatible change, so we'd probably need to make it optional (or deprecate the current behavior).
Wdyt?