Harnessing Discrete Representations For Continual Reinforcement Learning