Introduce nightly run end-to-end testing

[cloderic]

Objective

• Have the ability to detect regressions by periodically launch run and compare result with reference value

[cloderic]

What's done

Works on a few run scenario

1. Launch the scenario (similar to cogment run start_run)
2. Have a process that retrieve the total reward from mlflow regularly
3. When the target total reward is reached at least once

Remaining work

• Base working, to be integrated (issue with docker-compose version at the moment)

[cloderic]

To be considered -> Starting with snapshot of previously trained models?