Boosting LLM-based Relevance Modeling with Distribution-Aware Robust Learning