Improved Operator Learning by Orthogonal Attention