Optimal Regret Bounds for Collaborative Learning in Bandits