RFC/WIP Feature names within fit

[amueller]

This is an alternative to #12627 that I proposed in scikit-learn/enhancement_proposals#18

Basically I think now it's better to have the feature names as close to X as possible, so they are not out of sync, and I want the user interface to be as small as possible.

This PR adds feature_names_in as a parameter to fit, and adds feature_names_in_ as an attribute to every estimator, and feature_names_out_ as an attribute to all transformers.

Other alternatives that do basically the same but don't require an attribute to fit are:

• require the user to set the feature_names_in attribute manually
• pass around objects that have feature names attached to X, i.e. use dataframes or a dataset object or a subclass of ndarray that adds feature names.

[amueller]

this lets you trace the change of features within a pipeline / column transformer quite nicely

[jnothman]

What do you think of using some ugly naming convention, like 'ends with a _' to indicate fit parameters that are not sample-aligned. I.e. fit(self, X, y, feature_names_) and ``partial_fit(self, X, y, classes_, feature_names_)`

[amueller]

closing this, I think I prefer passing around a different datastructure if we want to have feature names in fit. A simpler first fix is to do get_feature_names and not have them during fit. See #18444 for the most up-to-date solution on this issue.