StructFormer: Document Structure-based Masked Attention and its Impact on Language Model Pre-Training