Provide relatively small model for ViT

[Gy-Lu]

Describe the feature

Hi, I find that we have provided too many huge models as examples, for instance, we reshape cifar-10 to 224*224 and use ViT Huge(at least not tiny) to train.
This reshape is quite redundant and so as ViT Huge for cifar-10.