From Optimality to Robustness: Dirichlet Sampling Strategies in Stochastic Bandits