Near-Optimal Collaborative Learning in Bandits