Enable automated model list copying for localized READMEs

[qqaatw]

What does this PR do?

Currently, the model list in each localized README such as README_zh-hans.md is updated manually, this PR introduces automated model list copying for localized READMEs. A proper tester for this change has been included in this PR.

The model list of a localized README is updated through the following steps:

1. Check if every model in the model list of README.md exists in a localized README, according to the model name e.g. BERT.
2. If a model doesn't exist in a localized README, the metadata of the model is fetched from README.md and substituted to a predefined localized format string.
3. Repeat step 1 and 2 until all models are checked.
4. Sort models in a localized model list alphabetically.

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@sgugger

[sgugger]

Thanks a lot for your PR, it looks very cool! I have left a few comments and let's also see if @JetRunner has some comments.

[JetRunner]

@qqaatw Thanks for taking care of this - I'm thinking since currently, the traditional Chinese version of readme doesn't translate this list, maybe we can just leave it untranslated for simplified Chinese as well. Does that simplify things?

[qqaatw]

@JetRunner Thanks for your reply!

I think this depends on the preference of simplified Chinese users. Translated or not has no difference since this PR introduces a way that directly captures metadata from the English version, and the captured metadata can be substituted into any predefined localized model description format string. Considering If there are more localized READMEs translated in the future, let's say Japanese version of README, this method can also apply to the model list of these READMEs.

In addition, some of models have supplemental data that can be manually translated after automated copying. For example, the text below is the supplemental data of DistilBERT:

The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.

The simplified Chinese version has the supplemental data translated, which looks good to me:

同样的方法也应用于压缩 GPT-2 到 [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation), RoBERTa 到 [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT 到 [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) 和德语版 DistilBERT。

[JetRunner]

@JetRunner Thanks for your reply!

I think this depends on the preference of simplified Chinese users. Translated or not has no difference since this PR introduces a way that directly captures metadata from the English version, and the captured metadata can be substituted into any predefined localized model description format string. Considering If there are more localized READMEs translated in the future, let's say Japanese version of README, this method can also apply to the model list of these READMEs.

In addition, some of models have supplemental data that can be manually translated after automated copying. For example, the text below is the supplemental data of DistilBERT:

The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.

The simplified Chinese version has the supplemental data translated, which looks good to me:

同样的方法也应用于压缩 GPT-2 到 [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation), RoBERTa 到 [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT 到 [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) 和德语版 DistilBERT。

Sounds great ;)

[qqaatw]

One exception I found is that there are few models (GPT-J, GPT-Neo, and T5v1.1) using released in the repository not the regular released with the paper. The localized description of this kind of models should manually be updated.

[qqaatw]

@sgugger The suggestions you provided have been applied. Thanks for the review!

[sgugger]

Just committed one last ypo fix, thanks a lot for your PR!