Remove words from DQA output

[Wauplin]

Related to (unfortunately) private slack convo (here).

originally from @Rocketknight1 :

Hey, for document-question-answering, the JS spec has "words", which is:
The index of each word/box pair that is in the answer
In transformers we don't output this at all. I'm not really sure how it's generated/used in the spec, since transformers already had start, end, answer, etc.

and then

Update: I think the words output never exists because of a pipeline bug. The pipeline has two code paths in several functions, one for VisionEncoderDecoder models and one for everything else.
In preprocess(), the VisionEncoderDecoder path always sets words to None which means it cannot be passed through to the output.
In postprocess() , the non-VisionEncoderDecoder path calls postprocess_extractive_qa. However, this function rewrites the answers dict without a words key
In other words, the preprocess() method deletes words for VisionEncoderDecoder models and the postprocess() method deletes it for everything else, so it always gets deleted! The right solution is just to remove it from the docstring and the JS spec.

---

Also to mention: DQA is only served from transformers in the Inference API.

[Wauplin]

(closing as duplicate of #1040)