Separate calculation of ensemble from prediction

[Martin O'Leary]

At the moment, the ensemble is calculated at prediction time. This is obviously inefficient if more than one predict() call is made. It also means that it's impossible to interrogate the ensemble object to determine which models were used, etc.