Better predict api design

[lolipopshock]

This PR introduces two improvements in the predict apis:

1. It can specify the return types of the predict API -- either layoutparser Layout or just a list of category predictions. This aims to make the API more generalizable and can support downstream uses like mmda

2. It adds the predict_page API that is dedicated for the vila datamodels. The prediction process is further simplified into one line:

   for idx, page_token in enumerate(page_tokens):
   
   # New
   predicted_tokens1 = pdf_predictor.predict_page(
       page_token, page_image=page_images[idx], visual_group_detector=vision_model
   )
   
   # Previous
   blocks = vision_model.detect(page_images[idx])
   page_token.annotate(blocks=blocks)
   pdf_data = page_token.to_pagedata().to_dict()
   predicted_tokens2 = pdf_predictor.predict(pdf_data, page_token.page_size)
   
   assert predicted_tokens1 == predicted_tokens2