Scalable Bayesian Uncertainty Quantification for Neural Network Potentials: Promise and Pitfalls