Verifiable Reinforcement Learning Systems via Compositionality