Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation