Serving Heterogeneous LoRA Adapters in Distributed LLM Inference Systems