Corrected Hanabi, new Dockerfile, python 3.10 and other fixes

[mttga]

This merge contains the following updates:

• A new, properly tested Hanabi environment from which we can achieve SOTA performances (+ nice features like rendering, playing manual games, and the possibility to use pretrained OBL models).
• A new Dockerfile based on the NVIDIA official JAX image.
• A refinement of the requirements files, avoiding multiple installation types (dev, qlearning, etc.). Also, the requirements become very strict in order to mirror as much as possible the NVIDIA image and to prevent people from using different JAX-JAXlib-FLAX-BRAX versions (important since we're re-collecting results in the next sprint). This might be too strict in some cases, but too many people reported installation issues, so I think it's worth it.
• Support for Python 3.10 (and removal of support for Python 3.8-3.9). This is also done with the goal of mirroring the NVIDIA image, which uses Python 3.10.
• A fix for issue #66 (which also drastically improves the reported TransfQMix results on the "sz" based maps).
• Fix of a deprecated use of jnp.concat (instead of jnp.concatenate) in smax, which causes problems when using the NVIDIA image.

[mttga]

It says 10k additions which is crazy but it's because we're adding a ground-truth file for testing hanabi which contains 10k game scores.

[benellis3]

Have you tested that the PPO code runs with the new requirements for other environments? Otherwise this could break a lot of stuff.

[mttga]

Have you tested that the PPO code runs with the new requirements for other environments? Otherwise this could break a lot of stuff.

I've checked that all the scripts run and I collected results for both IPPO and MAPPO with rnns in mpe and smax here: https://wandb.ai/mttga/jaxmarl_pull_request_71?nw=nwusermttga

Let me know if you want to see other results

[mttga]

Overcooked:

[benellis3]

LGTM. Thanks very much for all this hard work this is awesome 😄 . Feel free to merge when you are ready.