Understanding Approximate Fisher Information for Fast Convergence of Natural Gradient Descent in Wide Neural Networks