Unifying design-based and model-based sampling inference by estimating a joint population distribution for weights and outcomes ∗