Relearning Forgotten Knowledge: on Forgetting, Overfit and Training-Free Ensembles of DNNs