Optimality of Thompson Sampling with Noninformative Priors for Pareto Bandits