A Simple Approach for State-Action Abstraction using a Learned MDP Homomorphism