Provably Efficient Q-learning for Robust Reinforcement Learning