Reconciling meta-learning and continual learning with online mixtures of tasks