hook new weights

[CarloLucibello]

also made some doc changes:

• fixed the links in the readme
• added to the readme table some missing model
• now docs/make.jl copies the readme to index.md so that the two don't fall out of sync

Fix #234, fix #206, fix #220, partially address #187

[darsnack]

One issue I missed in the last PR is the use of the main branch URL from HF. These should be updated to reflect a specific commit so that each release's artifacts continue to work as HF is updated.

[CarloLucibello]

One issue I missed in the last PR is the use of the main branch URL from HF. These should be updated to reflect a specific commit so that each release's artifacts continue to work as HF is updated.

will do, although we should avoid changing files once they are used in a tagged release. Updated weights should become a new hosted file and artifact.

[CarloLucibello]

ok this is it for this PR, I'm sorry it got pretty big. I'll fix the residual problems in more focused PRs.

After merging this:

• for every model for which we had pretrained weights in the last tagged release we will continue to have pretrained weights, either in the old "save the whole model" format (squeezenet, densenet) or the new "save only the state" format
• we also have weights for the ViT, although the output has some slight discrepancies with pytorch

[darsnack]

Based on the test error, does the ResNeXt-101-64x4d SHA need to be updated?

[CarloLucibello]

turns out resnext_64x4 has only the IMAGENET1K_V1 weights

[CarloLucibello]

greenlights!

[darsnack]

Nice!