On the Approximation of Bi-Lipschitz Maps by Invertible Neural Networks