Role Prompting Guided Domain Adaptation with General Capability Preserve for Large Language Models