A Symmetry-Aware Exploration of Bayesian Neural Network Posteriors