Option for save_pretrained() to Export Model Source Code Files

### Feature request Add an optional flag (e.g., include_source_code=True) to PreTrainedModel.save_pretrained() / PreTrainedTokenizer.save_pretrained() that copies the currently-used Python implementation files (e.g., modeling_xxx.py, configuration_xxx.py, tokenization_xxx.py) into the target directory—even when those files already exist in the Transformers package. ### Motivation - **Editable snapshots** When iterating on research ideas, I often need to tweak the model’s forward pass or config class after fine-tuning. Editing code inside site-packages/transformers is brittle (breaks with upgrades, affects unrelated projects, requires virtual-env gymnastics). - **Reproducibility & archiving** Archiving both weights and exact code in a single artifact dramatically simplifies sharing and long-term reproducibility, especially when the upstream library may evolve. ### Your contribution I’m not very familiar with the internal mechanism Transformers uses to store model files; if the feature is feasible, I’m willing to submit a PR myself later.