feat(forecast): Rust forecast crate, bit-for-bit with the Python serving path

Phase 2a of docs/RUST_REWRITE.md. rust/forecast (ping-forecast) ports the
serving half of src/ml: the HII mirror gap fill, the hourly grid, every
feature (pandas' Kahan rolling sums included), the Mae Ngat dam frame,
HistGradientBoosting tree inference, the heuristic fallback, stale-threshold
handling and the P.1 flood-stage probabilities.

Hand-over from Python (src/ml/export.py):
- models/flood_<station>.json.gz next to each joblib bundle; train_all() writes
  it, retrain.sh promotes it, `python -m src.ml.export` exports existing ones.
- rust/forecast/flood_config.json: thresholds, stages, leads and notify
  constants, compiled in. tests/test_model_export.py fails while it differs
  from the Python source.

Parity (rust/forecast/tests/parity.rs, fixtures from
testdata/gen_forecast_fixtures.py on the production code path): four
scenarios incl. the 2024 flood with RID lagging behind HII. Merged readings,
all features and all levels identical; with all 15 local models 377/384
probabilities bit-identical, the rest within 2 ulp (exp).

Python refactor without behaviour change: predict.forecasts_from_long, and
get_latest_forecasts/load_latest/serving_frame take now/rain for
reproducible comparisons; the HII query orders its rows.

src/ml/train.py

@@ -554,6 +554,10 @@ def train_all(
                 station_results[station] = station_metrics
                 continue
             joblib.dump(bundle, models_dir / f"flood_{station}.joblib")
+            # the Rust forecaster's copy (src/ml/export.py)
+            from .export import write_bundle_json
+
+            write_bundle_json(bundle, models_dir / f"flood_{station}.json.gz")
             logger.info(
                 f"{station}: trained, {bundle['n_train_rows']} rows, "
                 f"{len(bundle['skipped_heads'])} heads skipped"