Counterfactual Augmentation for Multimodal Learning Under Presentation Bias