Plug-and-Play Performance Estimation for LLM Services without Relying on Labeled Data