support running via Cog

[bghira]

This pull request adds Cog support for training with SimpleTuner by introducing a new Predictor entrypoint, a helper orchestration layer, and configuration files. These changes enable users to launch training jobs through Cog, handling dataset staging, configuration, and output packaging automatically.

The most important changes are:

Cog integration and configuration:

• Added a new cog.yaml file with build settings (CUDA version, Python version, system packages) and specified the predictor entrypoint as predict.py:Predictor.
• Created a comprehensive .dockerignore to exclude unnecessary files from the Docker build context, improving build performance and security.

Predictor and orchestration logic:

• Added predict.py, which defines a Predictor class for Cog, handling user inputs, launching SimpleTuner training jobs, and returning zipped output directories.
• Introduced simpletuner/cog.py, a utility module that stages datasets, prepares config files, runs training, and packages results for Cog, including safe archive extraction and Hugging Face token handling.

[ParagEkbote]

Could we push the trained model checkpoint or model to Replicate?

cc: @bghira

[bghira]

@ParagEkbote see this doc