Efficient MCMC Sampling for Bayesian Matrix Factorization by Breaking Posterior Symmetries