PertEval: Unveiling Real Knowledge Capacity of LLMs with Knowledge-Invariant Perturbations