Bayesian decision-making under misspecified priors with applications to meta-learning