Flexible Bayesian Last Layer Models Using Implicit Priors and Diffusion Posterior Sampling