Meta-Learning and Universality: Deep Representations and Gradient Descent can Approximate any Learning Algorithm