Discrete Factorial Representations as an Abstraction for Goal Conditioned Reinforcement Learning