On the Importance of Strong Baselines in Bayesian Deep Learning