Fix DistilBERT prediction output shape and model loading configuration

[Irozuku]

Summary

Fixes issues in the DistilBERT transformer model affecting binary classification predictions and model loading. Ensures prediction probabilities have a consistent shape and restores missing logging parameters when loading a model, preventing misconfiguration.

---

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)

Briefly list the important modified files and what was done.

• distilbert_transformer.py:
  • Fixed probability collection in predict by appending batch outputs and stacking with np.vstack to ensure consistent shapes.
  • Fixed load method to properly restore logging-related parameters, preventing loss of configuration when reloading models.

---

Testing (optional)

• Verify that training the model runs successfully without errors.
• Verify that the trained model can be used to generate predictions correctly (especially for binary classification).