How Regularization Terms Make Invertible Neural Networks Bayesian Point Estimators