Enhancing a large language model with a chain-of-metacognitive reasoning approach increases argumentative writing evaluation accuracy, student writing outcomes, and mental effort