Safety Verification of Autonomous Systems: A Multi-Fidelity Reinforcement Learning Approach