Predictive posterior sampling from non-stationnary Gaussian process priors via Diffusion models with application to climate data