Synthesis And Verification Of Neural Control Barrier Functions For Safe Reinforcement Learning With Guarantees