Vision Language Models Map Logos to Text via Semantic Entanglement in the Visual Projector