Thresholding Bandit with Optimal Aggregate Regret