Probabilistic Guarantees for Safe Deep Reinforcement Learning