Improving Neural Additive Models with Bayesian Principles