LLM-based Rewriting of Inappropriate Argumentation using Reinforcement Learning from Machine Feedback