On the Surprising Effectiveness of Attention Transfer for Vision Transformers