Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned