Don't Learn, Ground: A Case for Natural Language Inference with Visual Grounding