[ENH] handling of `pd.DataFrames` in importance estimators

[jpaillard]

sklearn estimators handle pandas' dataframes well.
However, our importance methods convert all dataframes to arrays, leading to warnings "estimator was fitted on named features". It would be nice to remove such warnings by improving the way we handle dataframes.

[bthirion]

For the 0.4 release ? Or do you believe we need it urgently ?

[jpaillard]

IMO it's a nice to have but nothing urgent