Probabilistic Constraint for Safety-Critical Reinforcement Learning