ReflectRM: Boosting Generative Reward Models via Self-Reflection within a Unified Judgment Framework