Test model training

[RunDevelopment]

This adds a util function to test that models can train and uses it to test all image models.

Out of 44 models, 10 fail. I added a TODO comment to all of those. We have to decide how to deal with them.

---

I also noticed that it's impossible to use the call API for training, which is a shame.

[joeyballentine]

I don't think the call API should be used for training anyway imo. People training the arches should be using the raw output without any of the extra things we do like clipping, since they wouldn't be running losses on the actual model output

[joeyballentine]

Alternatively, we could have an optional named arg for whether or not the call api is being used during training, and then not do eval if thats set to true?

[joeyballentine]

But, i still kinda think that if you're training, you shouldnt be using our model descriptor and should instead just be loading the raw arches