Rectified SpaAttn: Revisiting Attention Sparsity for Efficient Video Generation