Prover-Verifier Games improve legibility of language model outputs