[refactor] Move output files to a common results/ folder

[ervteng]

Proposed change(s)

This PR changes the artifact output directory to this structure:

results/ 
    {run_id}/
        configuration.yaml - the configuration (after CLI overrides) used for this run.
        {behavior_name}/
            Tensorboard files
            snapshot.ckpt (and .pb) - model checkpoints and intermediate output files
        {behavior_name}.csv - CSV log
        {behavior_name}.nn - Barracuda Output
    run_logs/
        timers.json - timers file

In addition, we now output the RunOptions object as a YAML file configuration.yaml. When combined with PR #3815, the resulting YAML can then be used with mlagents-learn to re-run the same run again.

Useful links (Github issues, JIRA tickets, ML-Agents forum threads etc.)

https://docs.google.com/document/d/1uyk5JVNevfWy2DmqVHy19lcDzyQCNppPxkPOgJcQ7z4/edit#heading=h.5cr9b5kpfi98

Types of change(s)

• Bug fix
• New feature
• Code refactor
• Breaking change
• Documentation update
• Other (please describe)

Checklist

• Added tests that prove my fix is effective or that my feature works
• Updated the changelog (if applicable)
• Updated the documentation (if applicable)
• Updated the migration guide (if applicable)

Other comments

[ervteng]

@anupambhatnagar - Added you to the review for visibility as it will require a corresponding change to the CI pipeline.

[harperj]

@ervteng you're right that this will break the daily CI runs. I think there are 3 places in the CI we'll need to update:

• results reporter
• script for updating NN files
• the runner script's barracuda inference logic

One of us could take a stab if that would be useful.

Unrelatedly, I think the structure you listed in the PR description is a little off -- all files should be nested under the run-id folder. I'd also suggest we might just get rid of the run_logs folder and move those files into the general run-id folder.

[ervteng]

@harperj I've actually already made the CI changes in this PR: https://github.com/Unity-Technologies/ml-agents-cloud/pull/105 but definitely it needs a looksee from one of you to make sure I didn't miss anything. Seems to work when I tested it.

I think getting rid of the run-log (summaries) in the design doc could work, as there will only be two files in there (the timers.json and eventually the Player.Log). The only time it will make sense to have a separate run-log is when running many environments, as each will have their own Player.log (so Player-1.log, Player-2.log, etc.) and clutter up the {run-id} directory.