TODO: motivate freezing: pipeline components, calibration, transfer/semisupervised learning

This should probably be a SLEP, but I just want it saved somewhere.

Features required for estimator freezing:

• clone must have is isinstance(obj, FrozenModel): return obj (or do so via class polymorphism / singledispatch)
• FrozenModel delegates all attribute access (get, set, del) to its wrapped estimator (except where specified)
  • hence its estimator cannot be accessible at FrozenModel().estimator but at some more munged name.
• FrozenModel has def fit(self, *args, **kwargs): return self
• FrozenModel has def fit_transform(self, *args, **kwargs): return fit(self, *args, **kwargs).transform(self, args[0]) (and similar for fit_predict?)
• isinstance(freeze(obj), type(obj)) == True and isinstance(freeze(obj), FrozenModel) == True
  • since this is determined from type(freeze(obj)) (excluding __instancecheck__, which seems irrelevant), this appears to be the hardest criterion to fulfill
  • seems to entail use of a mixin, class created in closure, setting __class__ (!), overloading of __reduce__, help! I think I've gone down the wrong path!!
• must behave nicely with pickle and copy.[deep]copy
• freeze(some_list) will freeze every element of the list