DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models