NLoRA: Nystr\"om-Initiated Low-Rank Adaptation for Large Language Models