vLLM-Omni: Fully Disaggregated Serving for Any-to-Any Multimodal Models