Preventing Gradient Attenuation in Lipschitz Constrained Convolutional Networks