RRPO: Robust Reward Policy Optimization for LLM-based Emotional TTS