Semantic Composition in Visually Grounded Language Models