Revealing the structure of language model capabilities