On the Complexity of Offline Reinforcement Learning with $Q^\star$-Approximation and Partial Coverage