Bayesian credibility for GLMs