The structure of the token space for large language models