Add stopping criteria to the scratch implementation

Extend the scratch decision tree to include max_depth and min_samples_split parameters to prevent infinite recursion and overfitting. Train models with two different depth limits and compare their train vs. test metrics.

Locate the scratch decision-tree implementation or notebook and read the model training flow first. Add the requested max_depth and min_samples_split controls, then train models with two depth limits and compare their train and test metrics. Done means recursion is bounded and the metric comparison is shown.

Written by the indexing model from the issue text.