Implement EfficientNet

[darsnack]

Follow-up from #106 (closes #106).

[theabhirath]

Is there also a way we can port the weights functionality from #113? Would be a good complement to #164

[darsnack]

Probably using the same technique as #164. I think any model that is plain convolutions, batch norm, and fully connected can be done in this way. There a gist linked on the HF repo model cards if you want to tackle adding this for other models in the repo.

[darsnack]

Test errors are happening later for the same memory issues as always. I've already limited the tests to the smallest subset that seems necessary.

[theabhirath]

Seems all good, but I'm worried that EfficientNets, which are the least heavy of the models, are causing this....or maybe it's a cumulative effect from the other models (Is GC.gc() perfect?)

[darsnack]

The tests get passed EfficientNet and error around ConvNeXt. Maybe ConvNeXt is right on the threshold for the CI?

[theabhirath]

Might be the MLPMixers, since those are quite large (and we are testing for the :large models there). But just a hunch

[theabhirath]

Seems like it was the MLPMixers after all, nice!

[theabhirath]

Could maybe try not testing for the ConvNeXt and ConvMixer large models, but I'm not sure if those are the only pain points for memory here - Xception and InceptionResNetv2 were pretty bad as well