XGBoost model training script on IEEE-CIS Fraud Detection dataset

[Teboho66]

Description

SentinelPay's MLModel entity and MLModelService are fully implemented
but currently use simulated inference. This issue trains a real XGBoost
model on the IEEE-CIS Fraud Detection dataset and registers it in MLflow.

Tasks

• Download IEEE-CIS dataset from Kaggle
• Create ml/train_xgboost.py with feature engineering matching
  FeatureVector schema
• Train XGBoost with: precision ≥ 0.85, recall ≥ 0.80 (BR-ML1)
• Register model in local MLflow: mlflow models serve
• Update MLModel.score() to call real inference

Acceptance Criteria

• Model achieves precision ≥ 0.85 AND recall ≥ 0.80 on holdout set
• POST /api/ml-models/{id}/evaluate returns meets_promotion_gate: true
• Model artifact path registered in MLflow