On the Expressivity of Neural Networks for Deep Reinforcement Learning