Handle sklearn version difference in ROCAUC scoring

[Creylay]

Summary

Older sklearn versions raise ValueError in roc_auc_score when a class is absent from a split during One-vs Rest (OvR) evaluation, while newer versions (e.g. 1.7.2) return NaN with a warning instead. This caused training jobs to crash on machines with older sklearn while working fine on machines with newer sklearn. The fix wraps the roc_auc_score call in a try/except ValueError so both behaviors normalize to float("nan"), which base_model.calculate_metrics already filters out via math.isfinite.

---

Type of Change

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

---

Changes (by file)

• DashAI/back/metrics/classification/roc_auc.py: wrap roc_auc_score call in try/except ValueError to return float("nan") instead of propagating the exception when a class is absent from a split in OvR mode.

---

Testing (optional)

Train a classification model with an imbalanced dataset (no stratified split) so at least one class is absent from validation/test. Confirm training completes and ROCAUC is omitted from those splits instead of crashing the job.

---

Notes (optional)

This bug only manifests when: (1) the task is multiclass, (2) the dataset is split without stratification, and (3) at least one class has no samples in the validation or test split. The calculate_metrics call happens inside the training loop (per epoch/step), so the crash occurs early during training rather than at the end.