Accelerating Unbiased LLM Evaluation via Synthetic Feedback