Continual Learning as Computationally Constrained Reinforcement Learning