[feat] `response_format` differentiation

[pkarw]

Additionally, for docling, I think it would be good to add something like a response_format parameter. Then:

• default – returns text, as it does now,
• object – returns the Docling object,
• meta_data – returns document metadata (such as the title inferred by the LLM and other details),
• bbox – returns bounding boxes (an object containing bounding box data), which will work with easyOCR.