Docs: improve Reader docs (models)

[anakin87]

(stems from #5340)

https://docs.haystack.deepset.ai/docs/reader

• There is not a clear distinction between base models (to be fine-tuned) vs pretrained models (which can just be used out-of-the-box)
• similar: no clear separation between using Reader to perform Question Answering and using it to fine-tune base models
• The proposed pretrained models are outdated (we are not promoting our most recent QA models)

(Don't mean to be harsh, I'm just in a hurry. 😃 )

[anakin87]

Related Discord message: https://discord.com/channels/993534733298450452/1130211893508853861/1131319426180984863

To understand which the better Reader models are, we can ask Sol-Eng...

[anakin87]

These benchmarks can also help: https://paperswithcode.com/sota/question-answering-on-squad-v2

[anakin87]

Although it is not urgent, I think that providing some suggestions to users on models to use is still a good idea for 2.0.

This page of DC docs may serve as inspiration.