Verifiable and Compositional Reinforcement Learning Systems