Adaptive Guidance Semantically Enhanced via Multimodal LLM for Edge-Cloud Object Detection