Computational-Statistical Gaps in Reinforcement Learning