Establish review process for training code

[maanug-nv]

Need to determine how contributions to megatron.training will be reviewed. Do we reuse the same expert groups? Do we need new final reviewers?

[Saad-Mallebhari]

@maanug-nv
For most changes touching megatron/training/, the existing expert groups should be sufficient , parallelism reviewers for anything affecting forward_backward, optimizer experts for gradient/LR changes, etc.
Where I'd suggest a new dedicated reviewer is the core training loop (training.py, initialize.py) , this doesn't cleanly fall under any existing group and is high enough risk that it warrants 1-2 designated final approvers + a CODEOWNERS entry to auto-assign them on PRs
A lightweight checklist for training PRs would also help reviewers move faster without missing regressions (parallelism configs, checkpoint compatibility, MFU impact).