On permutation symmetries in Bayesian neural network posteriors: a variational perspective