frameworks: improve log_artifact for model "checkpoints"

[dberenbaum]

When auto-logging model artifacts as part of the frameworks in #500, sometimes frameworks save multiple model files at different paths within a directory to allow users to keep multiple model checkpoints:

• https://lightning.ai/docs/pytorch/stable/common/checkpointing_basic.html
• https://huggingface.co/docs/transformers/main_classes/trainer#transformers.Trainer.save_model
• https://docs.fast.ai/callback.tracker.html#savemodelcallback

This present some questions for our callbacks:

1. Should we save the individual files or the entire directory?
2. How should we keep track of the latest model file so that it can be used to resume/analyze/deploy?

[dberenbaum]

After #472, live.log_artifact should be adding something to dvclive/dvc.yaml like:

artifacts:
  fast-ai-model:
    path: model.pth
    type: model

This artifacts section could be used to answer question 2 above. Each call to live.log_artifact can update the path in that artifacts section from model_1.pth to model_2.pth and so on, and @aguschin is working on making it possible to do things like dvc get . fast-ai-model so that it's possible to get the model without specifying the exact path of the last checkpoint.

In that case, question 1 becomes less important and is mostly about how we want to organize the .dvc files: is it okay to have one for each checkpoint, or should we try to track the model directory?

[dberenbaum]

I think we can close this as completed with the work we have done so far in lightning and huggingface