The Limitations of Large Width in Neural Networks: A Deep Gaussian Process Perspective