LoRA-Switch: Boosting the Efficiency of Dynamic LLM Adapters via System-Algorithm Co-design | Synapse