Are Layout-Infused Language Models Robust to Layout Distribution Shifts? A Case Study with Scientific Documents