Multimodal Neurons in Pretrained Text-Only Transformers