New feature: structured logs in rlberry

[TimotheeMathieu]

I think the outputs in rlberry are not always super nice. It would be interesting to use structured logs, for example dumping the rewards and various metrics in a text file like json or recfile that can be queried during training would be nice and cleaner than the output that we have for now. Also, it would allow us to have plotting tools that could create plots of reward during training if needed.

I plan to do a PR proof of concept of this soon-ish.