Can Transformers Learn Full Bayesian Inference in Context?