Add SENet Blocks in Encoding Layers

# 🚀 Feature Request

I read the article "[SesameBERT: Attention for Anywhere](https://arxiv.org/pdf/1910.03176.pdf)" and would like to add SENet blocks in the Huggingface implementation. The article's authors made an implementation with [Tensorflow](https://github.com/ICLR2020Sesame/SesameBert/blob/master/modeling.py), but I would like to use the lib in pytorch.

## Motivation

The use of ([Squeeze-and-Excitation Networks](https://arxiv.org/abs/1709.01507)) SENet Blocks has obtained state-of-the-art results. And they seem to be promising in NLP.

## Your contribution

I know that it is possible to modify the [[BertLayer()](https://github.com/huggingface/transformers/blob/61c506349134db0a0a2fd6fb2eff8e29a2f84e79/src/transformers/models/bert/modeling_bert.py#L430)] and [[BertEnconder()](https://github.com/huggingface/transformers/blob/61c506349134db0a0a2fd6fb2eff8e29a2f84e79/src/transformers/models/bert/modeling_bert.py#L513)] classes

Any suggestions on how to modify the code so that you can apply the idea used in the article?