Holmes: Benchmark the Linguistic Competence of Language Models