Auditing Prompt Caching in Language Model APIs