Implement Explainable Boosting Machine (EBM) pricing model #43

Problem

We chose the Explainable Boosting Machine (EBM) as our pricing model, but it isn't implemented
yet. The recommendation endpoint still returns a placeholder.

Goal

Train an EBM on booking history and external signals. Serve real recommendations through the
existing endpoint, each with a rate, confidence interval and explanation.

Scope

• Add interpret to pyproject.toml. Create services/ml/ebm.py, with EBMModel(Model) wrapping
  ExplainableBoostingRegressor.
• Add a feature builder shared by training and inference, with DB access via repositories
  only. Features: calendar, events, weather, flights, rate vs base rate.
• Add a training script with a time-based split that saves a model artifact. Put the artifact
  path in config.py.
• Wire the model into GET /hotels/{id}/recommendations. Return the rate, a conformal
  confidence interval and the top contributing terms as the explanation.
• Add EBM to compare_models.py.

Acceptance criteria

• Training script reports MAE/RMSE/R² and saves a loadable artifact.
• Endpoint returns a model-generated rate, interval and explanation in the agreed shape.
• Missing model or too little history returns a clear error, not a 500.
• No route or service touches the session directly.
• Tests cover the model, the feature builder and the endpoint (happy path + 404).