Add probabilistic multi-horizon forecasting with conformal calibration

[xungirl]

Motivation

• Provide backward-compatible probabilistic, multi-horizon forecasts (default pred_len=96) on top of existing point-forecast backbones.
• Enable calibrated interval outputs via split/joint conformal methods and optional online adaptation to meet coverage guarantees.
• Make the implementation modular and minimally invasive so existing point-only experiments continue to work when probabilistic features are disabled.

Description

• Added a model-agnostic forecasting package app/forecasting/ implementing ForecastConfig and ForecastPipeline to wrap existing backbones and expose flags for quantiles, conformal mode, and regularizers.
• Implemented a monotonic quantile head MonotonicQuantileHead with median-centered softplus increments to avoid quantile crossing and produce outputs shaped [B, H, D, Q].
• Implemented pinball loss, optional crossing penalty, horizon smoothness regularizer, and optional trajectory-level consistency loss under app/forecasting/losses.py.
• Added split conformal utilities SplitConformalCalibrator supporting horizon-wise and joint trajectory calibration, optional adaptive updates (rolling-window + EMA), and a coverage-drop reset hook in app/forecasting/conformal.py.
• Added probabilistic metrics (horizon/average coverage, mean interval width, Winkler, WIS approximation, joint trajectory coverage) in app/forecasting/metrics.py.
• Added an end-to-end demo scripts/forecasting_example.py, unit tests tests/test_forecasting_probabilistic.py, and README examples; and updated requirements (requirements.txt, requirements-backend.txt) to include numpy.
• Modified/added files: app/forecasting/__init__.py, app/forecasting/quantile.py, app/forecasting/losses.py, app/forecasting/conformal.py, app/forecasting/metrics.py, app/forecasting/pipeline.py, scripts/forecasting_example.py, tests/test_forecasting_probabilistic.py, README.md, requirements.txt, and requirements-backend.txt.

Testing

• Compiled the new modules with python -m compileall app/forecasting scripts/forecasting_example.py tests/test_forecasting_probabilistic.py which completed successfully.
• Ran pytest -q tests/test_forecasting_probabilistic.py initially, which failed during collection with ModuleNotFoundError: No module named 'numpy' in the environment before dependency update.
• Added numpy>=1.24.0 to requirements.txt and requirements-backend.txt and recompiled; unit tests cover monotonic quantile ordering, conformal horizon widening, and point-only backward compatibility (intended to pass once numpy is installed in the test environment).

---

Codex Task