ViT tensorflow Implementation

# 🚀 Feature request

I was reading about ViT in the HuggingFace document and noticed there is no TF implementation of it. It would be great to have it in HuggingFace repo.

## Motivation

I have seen [this](https://keras.io/examples/vision/image_classification_with_vision_transformer/) and think it wouldn't be so hard. We can convert pytorch pretrain weights and use it for tensroflow model.