Beyond Target Networks: Improving Deep Q-learning with Functional Regularization