DOPE: Doubly Optimistic and Pessimistic Exploration for Safe Reinforcement Learning