Offline Reinforcement Learning: Fundamental Barriers for Value Function Approximation