Support for Monotonic Mulithead Attention based Simultaneous Speech-to-text Translation

# 🌟 New model addition
Simultaneous Speech-to-text Translation using Monotonic Multihead Attention(MMA).
I am wondering if anybody is working on implementing this model for now.
However, I am worried that if this model is going to be supported by Hugging Face systems, since inference works in a particular way using frameworks like [SimulEval](https://github.com/facebookresearch/SimulEval) to simulate streaming input which may not be compatible with current Hugging Face's inference system?

## Model description
[MMA(Ma et al., 2019)](https://arxiv.org/abs/1909.12406) has been used to handle streaming text/speech inputs mostly for translation, where MMA extends the monotonic attention mechanism to multihead.

## Open source status
* [x] the model implementation is available: [Fairseq Implementation is available here](https://github.com/pytorch/fairseq/blob/fcca32258c8e8bcc9f9890bf4714fa2f96b6b3e1/examples/simultaneous_translation/models/convtransformer_simul_trans.py#L29~#L63)
* [ ] the model weights are available: (give details)
* [ ] who are the authors: (mention them, if possible by @gh-username) : Xutai Ma(@xutaima), Juan Pino, James Cross, Liezl Puzon, Jiatao Gu

Inference framework : [Facebook Research SimulEval](https://github.com/facebookresearch/SimulEval)

Start by reviewing the linked Fairseq implementation in examples/simultaneous_translation/models/convtransformer_simul_trans.py and the SimulEval inference framework. Clarify how streaming inference should fit Hugging Face systems, whether model weights will be available, and what acceptance criteria define support; the issue currently leaves these decisions unresolved.

Written by the indexing model from the issue text.