Improving Proactive Dialog Agents Using Socially-Aware Reinforcement Learning