Optimal approximate sampling from discrete probability distributions