KoLA: Carefully Benchmarking World Knowledge of Large Language Models