More Than Catastrophic Forgetting: Integrating General Capabilities For Domain-Specific LLMs