Approximation of laws of multinomial parameters by mixtures of Dirichlet distributions with applications to Bayesian inference