Model-free Safe Control for Zero-Violation Reinforcement Learning