Can Language Model Moderators Improve the Health of Online Discourse?