CLoQ: Enhancing Fine-Tuning of Quantized LLMs via Calibrated LoRA Initialization