Eigenvalue Distribution of Large Random Matrices Arising in Deep Neural Networks: Orthogonal Case