Rethinking the Evaluation for Conversational Recommendation in the Era of Large Language Models