Add LightGBM as third model candidate in regression pipeline

[agentksimha]

This PR adds LightGBMRegressor as a third competing model alongside RandomForest and XGBoost in the regression pipeline.

What changed:

• Added LGBMRegressor import
• Added param_grid_lgbm with learning_rate, n_estimators, max_depth, num_leaves
• Added LightGBM as a third entry in model_list so it participates in the same GridSearchCV and MSE-based selection logic

Justification:
LightGBM uses leaf-wise tree growth instead of depth-wise (XGBoost/RF), which is better suited for the feature-engineered dataset containing ratio and difference features with skewed distributions. It is faster to train and more memory-efficient on tabular data, making it a justified addition without disrupting the existing pipeline structure.

Results:
R² improved across all four non-special targets (CR21, CR54, CR13, CR23). Threshold accuracy showed mixed results — improved on CR21 and CR54, slightly lower on CR13 and CR23 — reflecting LightGBM's tradeoff of minimising squared error over threshold-based accuracy on those columns.

No changes to preprocessing, feature engineering, special targets, or output format.