Revert transformers path

[Irozuku]

Summary

Reverts a previous pull request that removed temporary checkpoint handling during model training. This restores the behavior where models can use a dedicated temporary directory for Hugging Face training checkpoints and ensures the directory exists during the build process.

---

Type of Change

Check all that apply like this [x]:

• Backend change
• Frontend change
• CI / Workflow change
• Build / Packaging change
• Bug fix
• Documentation

---

Changes (by file)

• .github/workflows/build.yml: Restores creation and inclusion of the DashAI/back/user_models/temp_checkpoints directory in PyInstaller build steps for all platforms.
• DashAI/back/models/.../distilbert_model.py: Restores the use of a temporary checkpoint directory for the Hugging Face Trainer and cleanup of the directory after training.
• DashAI/back/models/.../opus_mt_en_es_transformer.py: Reintroduces the temporary checkpoint directory for training and removes it after training completes.

---

Testing (optional)

• Verify that the build workflows correctly create and package the temp_checkpoints directory.
• Confirm that model training runs successfully and that temporary checkpoints are cleaned up after training.

---

Notes (optional)

This PR reverts the removal of temporary checkpoint handling, restoring support for models that require temporary storage during training while preventing leftover checkpoint files.

[cristian-tamblay]

I dont know if there's a way to parametrize this so windows+mac executables use the same flags (or at least some), because im sure we will fail adding these to both of them in the future

[cristian-tamblay]

Fix conflicts and maybe try to parametrize flags?