feat(notebooks): add 04_model_validation benchmarking notebook

[franchaise]

Summary

• Adds notebooks/04_model_validation.ipynb — validates segmentation predictions against reference masks and the biomass regressor against held-out labels across Amazon, Congo, and Southeast Asia.
• Computes per-tile IoU, F1, precision, recall, accuracy via a confusion-matrix helper, plus per-region regression metrics (RMSE / MAE / R^2 / MAPE).
• Aggregates per-region and mean values into outputs/validation/benchmark_report.json — the single artifact consumed by scripts/governance_ci_gate.py and governance.model_card.build_model_card.
• Falls back to synthetic tiles when real ground-truth data is not present, so the notebook stays runnable in CI.

Why

Sprint deliverable: "Write 04_model_validation.ipynb with baseline validation results."

Test plan

• Valid notebook JSON.
• Runs top-to-bottom on synthetic data via papermill without errors.

Notes for reviewers

• Imports validate_predictions from analytics.validation (already on develop) and BiomassRegressor from feat(models): add biomass and carbon-stock regression module #46.
• The benchmark report shape is documented in the closing cell so downstream consumers know what to expect.

[obielin]

Happy with this from the governance angle — the per-region table is exactly the shape build_model_card consumes for the Fairness section once we attach a region-level disparity metric on top.

One thought for a follow-up (not for this PR): the synthetic confusion matrix is built with a uniform 8% disagreement rate, so the per-region metrics look identical. When real ground-truth tiles land we'll see actual divergence. Approving.