On Reconstructing Training Data From Bayesian Posteriors and Trained Models