Document Classifier could use pre-trained embedding models

[bigabig]

Currently, the document classifier is trained on a general language model (e.g. modern-bert-base). However, similar to sentence classifiers, we could leverage pre-trained embedding models to reduce the training time!