Benchmarking Linguistic Diversity of Large Language Models