RevisEval: Improving LLM-as-a-Judge via Response-Adapted References