GPT, But Backwards: Exactly Inverting Language Model Outputs