WIP Add table summarizing classification weights accuracies in docs

[NicolasHug]

Rendered docs: https://output.circle-artifacts.com/output/job/e411a4b2-f155-422e-bd08-e03f854aecc6/artifacts/0/docs/models.html

Table looks like this

I'll make sure to add links from the weight names to their documentation page as in #5577
Ideally the rows should be more narrow but I don't know how to do that yet. Not too big of a deal anyway.

[datumbox]

@NicolasHug It looks pretty good. I assume we can fix the parameters to not use scientific format, right?

[NicolasHug]

I did it on purpose to give an idea of the order of magnitude directly, but I don't have a strong opinion, I'll revert

[NicolasHug]

Any thought on that @datumbox ?

[datumbox]

Yes you are right. We need a registration mechanism for the models. The new Datasets API has one, so part of the reason I didn't want to invest time creating one is to potentially adopt/extend the one on Datasets. Thoughts?

[NicolasHug]

The registration of the dataset is very very basic, it's just a decorator that adds the callable/object to a private dict. It would probably make sense to use something similar for the models / weights. Whether we should be relying on the same utils though is up for discussion - as a first version I'd suggest not to merge things and for the models to have a separate implementation. The code is really basic anyway.

[datumbox]

I'm happy with what you propose and on the technical details you mentioned. I had a similar simple approach on the original proposal of the Multiweights support but I didn't port it to adopt some solution in common with Datasets. The code doesn't have to be the same but I think the interface can be basic and similar. As discuss offline, the only thing different for models is the fact that there is a hierarchy (Detection, Optical Flow, Classification etc) and this needs to be taken into account because names across modules conflict (for example resnet50 exists both in Classification and Quantizaztion submodules).