A relatively common pattern is to change y according to some function for prediction, for example by using a logarithm in regression.
Usually you want to measure the performance in the original space, so you need to transform y back before scoring.
That's really easily done using a meta-estimator.
I think I reimplemented the logy meta-estimator about 10 times, so I think we should just add it. And we don't really need to restrict it to logarithm, we can allow the user to provide f and f_inverse.

Maybe should go into preprocessing?