Nexus:Proactive Intra-GPU Disaggregation of Prefill and Decode in LLM Serving