FedEval-LLM: Federated Evaluation of Large Language Models on Downstream Tasks with Collective Wisdom