Rethinking the Evaluation of Pre-trained Text-and-Layout Models from an Entity-Centric Perspective