Learning to Estimate Shapley Values with Vision Transformers