Number-specific tokenization changes

# 🌟 New model addition ## Model description I wanted to contribute a bunch of number-specific LMs proposed in recent work. Most of these are not architecture changes but simple tokenization tricks such as converting a number `329` to `3.29e2` (scientific; [Zhang et al. 2020](https://aclanthology.org/2020.findings-emnlp.439/)) or `3 2 9` (digit splitting; [Nogueira et al. 2021](https://arxiv.org/abs/2102.13019)) or `e2` (exponent only; [Spokoyny et al. 2020](https://aclanthology.org/2020.emnlp-main.385/) and [Thawani et al. 2021](https://aclanthology.org/2021.emnlp-main.557/)). The motivation is that several industrial applications require number-heavy NLP but struggle with existing models. I discussed a specific way to do this on team slack with @SaulLu for, say, NumBERT (scientific notation) which involves adding a new model (tokenizer-only) and uploading the pretrained weights to the hub. I wanted to open a broader discussion here about more such number-tokenizer-only methods, some of which may not even have pretrained weights. The hope would be to make some abstract intervention (perhaps at the tokenizer level) to let the user configure GPT or BERT tokenizer as `number_tokenizer=exponent`. But perhaps clubbing methods from different papers into one model/tokenizer is against HF's philosophy? If so, I could proceed with trying to simply incorporate them as individual models - is it fine if some of them do not have pretrained weights available? ## Open source status * [X] the model implementation is available: [T5 fine-tuned on arithmetic, based on HF transformers](https://github.com/castorini/transformers-arithmetic) and [NumBERT/scientific, based on google-bert original code](https://github.com/google-research/google-research/tree/master/numbert) * [X] the model weights are available: [BERT pretrained with scientific notation](https://console.cloud.google.com/storage/browser/gresearch/numbert) * [X] who are the authors: @spokoyny @XikunZhang @DeepakRamachandran…