Visual Document Understanding from Heterogeneous Documents Using Vision Transformers and Transfer Learning