Generalizing and Categorizing Skills in Reinforcement Learning Agents Using Partial Policy Homomorphisms