How much can language models memorize?