On the Limits of Innate Planning in Large Language Models