Fully Bayesian autoencoders with latent sparse Gaussian processes