Can you explain how the cyclic shift changes the feature map, and what position of the tokens is masked during the calculation of the attention? As in your paper's figure , it's too abstract for me...