Variational Inference Failures Under Model Symmetries: Permutation Invariant Posteriors for Bayesian Neural Networks