Mnemosyne: Learning to Train Transformers with Transformers