LowRA: Accurate and Efficient LoRA Fine-Tuning of LLMs under 2 Bits