Exponential Lower Bounds for Planning in MDPs With Linearly-Realizable Optimal Action-Value Functions