PEARL: Towards Permutation-Resilient LLMs