Bayesian hierarchical stacking: Some models are (somewhere) useful « Statistical Modeling, Causal Inference, and Social Science