Learning to Align Multi-Faceted Evaluation: A Unified and Robust Framework