feat: add SVD model with normalization, RMSE improved to 0.965

[harmandeep2993]

Summary

Significant model improvement by implementing SVD matrix factorization
with rating normalization, reducing RMSE from 2.52 to 0.965.

Changes

src/models/train.py

• Added train_svd_model() → SVD with configurable latent factors
• Added _evaluate_svd() → fast SVD evaluation
• Updated train_pipeline() → trains both ItemKNN and SVD
• Compares 6 models total via MLflow experiment tracking

src/models/evaluate.py

• Added calculate_rmse_svd() → full test set evaluation for SVD
• Added calculate_rmse_knn() → sampled evaluation for ItemKNN
• Updated evaluate_pipeline() → handles both model types

src/models/predict.py

• Added load_svd_artifacts() → loads SVD model + user means
• Added recommend_movies_svd() → instant SVD recommendations
• Updated predict_pipeline() → uses SVD, falls back to ItemKNN

src/features/build_features.py

• Added normalize_matrix() → subtracts user mean from ratings
• Added save_normalized_matrix() → saves normalized matrix + user means
• Updated build_features_pipeline() → returns normalized matrix

main.py

• Updated pipeline to pass normalized_matrix and user_means
• Clean up debug prints and unused imports

Model Comparison (MLflow)

Model	RMSE
ItemKNN K=10	2.72
ItemKNN K=20	2.82
ItemKNN K=50	2.88
SVD 50 factors	1.005 ← best ✅
SVD 100 factors	1.020
SVD 200 factors	1.023

Final Evaluation

• Full test set (199,708 rows)
• RMSE: 0.965 ✅
• Under 1.0 star error on 1-5 scale