Universality of invertible neural networks