SnapLLM: Switch between local LLM in under 1ms Multi-model&-modal serving engine