Learning World Graph Decompositions To Accelerate Reinforcement Learning