fix: correct fabricated TrainingTable methods in iterate-model skill

[9Olive]

skills/iterate-model.md recommended train_table.label_distribution() and train_table.count() (Step 4's data-quality table, and the Common Errors table) to diagnose class imbalance and row counts. Neither method exists on TrainingTable — confirmed against kumoai/pquery/training_table.py in kumo-sdk, whose only public methods are data_urls, data_df, export, and update. This is the same fabricated-method pattern already fixed in context/platform/sdk-overview.md by #80 (which established the correct .data_df()-based replacement pattern), just missed in this file.

Replaced with the pandas-based pattern already used elsewhere in this repo (and in #80): train_table.data_df()["target_col"].value_counts() for class-distribution and len(train_table.data_df()) for row counts.

3-line diff, no other changes.

[9Olive]

Closing: label_distribution() and count() are documented TrainingTable methods (sdk-overview.md:298-300; pitfall #7 at line 666 explicitly recommends them). Replacing them with data_df().value_counts() removes valid APIs and contradicts sdk-overview.md - nothing was broken here.

---

Automated internal-consistency review (cross-checked against the repo's own authoritative docs; a few SDK-surface claims spot-checked against installed kumoai 2.22.0).