Safe LoRA: the Silver Lining of Reducing Safety Risks when Fine-tuning Large Language Models