Describir: Relational non-verbatim unintended memorization in large language models