Particle algorithms for maximum likelihood training of latent variable models