Bandit Learning with Positive Externalities