An Inductive Synthesis Framework for Verifiable Reinforcement Learning