Compensating Distribution Drifts in Class-incremental Learning of Pre-trained Vision Transformers