feat(models): add biomass and carbon-stock regression module

[franchaise]

Summary

• Adds src/climatevision/models/regression.py — BiomassRegressor, a wrapper around sklearn RandomForestRegressor and xgboost.XGBRegressor with a stable fit / predict / evaluate / save / load API.
• Default feature ordering matches the spectral indices produced by the data preprocessor: NDVI, EVI, SAVI, NDMI, NBR, R, G, B, NIR, SWIR1.
• Helpers: biomass_to_carbon and biomass_to_co2e use IPCC defaults (carbon fraction 0.47, 44/12 ratio for CO2e).
• evaluate_regression computes RMSE / MAE / R^2 / MAPE for the eval and model-card pipelines.
• estimate_biomass_from_indices accepts a dict of per-pixel index arrays and runs inference in one call.
• Wires symbols into models/__init__.py.

Why

Sprint deliverable: "Build carbon.py — Random Forest & XGBoost regression for biomass prediction" / "Implement metrics for regression evaluation (RMSE, MAE, R-squared)." Backs the carbon analytics module Francis is delivering this sprint.

Test plan

• pytest tests/test_regression.py — 11/11 pass
  • biomass_to_carbon and biomass_to_co2e use the documented constants
  • evaluate_regression returns rmse=mae=0 and r2=1 on a perfect fit
  • shape-mismatch raises ValueError
  • RandomForest fits, predicts, and yields sensible metrics on synthetic data
  • feature_importances() returns one entry per feature name and sums to 1.0
  • unsupported model_type raises ValueError
  • predict() before fit() raises RuntimeError
  • save() / load() round-trip yields identical predictions
  • estimate_biomass_from_indices builds the right feature matrix and matches predict()
  • missing index in the dict raises KeyError
  • serialize_metrics writes a well-formed JSON

Notes for reviewers

• xgboost is a soft dependency — import xgboost is guarded so installs without it still work.
• models/__init__.py extends the existing __all__; no existing imports are touched.

[Goldokpa]

Solid foundation. The fit/predict/evaluate/save/load API matches what the analytics module expects and the test coverage (11 cases) hits the important branches — perfect-fit r2=1, save/load round-trip, missing-index KeyError. xgboost guarded as a soft dependency is the right call.

Minor nit (non-blocking): evaluate_regression returns NaN for r2 when ss_tot=0 — fine, just worth a short docstring note so callers don't trip on it. Approving.