Products of Many Large Random Matrices and Gradients in Deep Neural Networks