adopt the forecasting interface model for base abstract classifier so that rather than have abstract fit, have base fit call abstract _fit. This means that the validation of the input can be called in the base classifier.

However, for the classifiers this will require a reworking of capabilities, so that they are overridden and not shadowed. This is not a bad thing, but may well make it hard to roll capabilities into sklearn tags (which Im not sure I want to do in any case).

Tony