Initial Refactor and add new features for SAT solver including gpt_train.py and downstream tasks, plus addition of multi-node/multi-GPU processes

[cesposo]

This PR refactors key portions of the code, improves performance, and adds new features for SAT problem solving. Specific changes include:

• Extended/refactored gpt_train.py to include multiple model sizes, mostly for experiments with scale-laws
• Changed ancillary scripts downstream to gpt_train.py changes which includes alterations to tokenization previously made
• Added support for batch processing in folder_eval.py for Windows and Linux (eval_folder.py and eval_folder.sh)
• Improved logging and error handling during model evaluation.
• Addition of multi-node/multi-GPU batch processing for supercomputer cluster (experimental)

These changes aim to improve code maintainability and add new functionality for more flexible SAT problem solving.