How to integrate algo `Multi-Agent Transformer` in BenchMARL

Hi there,

First off, I want to say that BenchMARL looks like a fantastic resource for the MARL community! I've been following the project and I'm excited about the potential for benchmarking various algorithms.

I recently came across the Multi-Agent Transformer (MAT) by PKU-MARL, which is showcased in their GitHub [here](https://github.com/PKU-MARL/Multi-Agent-Transformer) and the accompanying paper [on arXiv](https://arxiv.org/abs/2205.14953). MAT introduces an innovative approach to MARL with its centralized joint-observation-action Transformer structure, and I believe it could be a valuable addition to BenchMARL.


I'm curious about how I might go about integrating MAT into BenchMARL. Could you provide some guidance or suggestions on this process? I'm eager to contribute to the project and explore the synergy between MAT and BenchMARL.

Thanks for your great work on this, and I'm looking forward to any advice you can offer!

The issue links to the Multi-Agent Transformer repository and paper but names no BenchMARL files, tests, or entry points. First review existing algorithm integrations and the MAT implementation to define the required scope; the work is done when MAT is integrated with appropriate benchmarks and its results are validated.

Written by the indexing model from the issue text.