Fix bug checkpoint IO bug once https://github.com/NVIDIA/NeMo/pull/1024 lands

[jstjohn]

Describe the bug
PR 10241 in NeMomakes a few improvements to checkpointing. In particular, it:

1. separates artifacts and checkpoint weights into separate directories
2. adds an option to only save model weights (skipping optimizer states) at the end of training (defaults to true)
3. removes save_best_model arg because it wasn't working anyway and having it present was misleading.

If there's demand for that arg on your end, we can add it to our list to support it properly. As of that PR save best checkpoint is getting dropped.

Make a change like the following in our checkpoint settings: NVIDIA/NeMo@10e4f88

This is tracked on the nemo side (internal to NVIDIA) with: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/issues/648

[jstjohn]

@skothenhill-nv FYI they're dropping save best since it's broken.