Changing categories and refit_final

[vsyrgkanis]

If one changes the categories and refits the final, nothing changes

est = LinearDML(discrete_treatment=True).fit(Y, T, X=X, cache_values=True)
est.categories = [0, 1]
est.refit_final()

The result doesn't use the categories at all. One could even do est.categories = [0, 1, 2] and it wouldn't throw an error.

Most probably refit_final should not be allowed if one changes the categories.

[kbattocchi]

At one point, we were considering an alternate approach to refitting where we kept track of which attributes were modified and prevented refitting if they would invalidate the results; however, we ultimately decided to go with the current approach, which puts the burden on the user to ensure that they only make changes that are compatible with refitting.