A Particle-Based Variational Approach to Bayesian Non-negative Matrix Factorization