Random Matrix Theory for Deep Learning: Beyond Eigenvalues of Linear Models