Index the per-season alpha in the seasonal ES models

[Venish Paneliya (VenishPaneliya)]

SeasonalExponentialSmoothing and SeasonalExponentialSmoothingOptimized fit one alpha per season, so model_["alpha"] has shape (season_length,). Two places were copied from the scalar SimpleExponentialSmoothing and use it unindexed, so both raise:

m = SeasonalExponentialSmoothing(season_length=12, alpha=0.1)
m.fit(y)

m.predict(h=6, level=[80])
# ValueError: operands could not be broadcast together with shapes (6,) (12,)

m.simulate(h=6, n_paths=4)
# ValueError: operands could not be broadcast together with shapes (12,) (4,)

Both variants are affected; the scalar SimpleExponentialSmoothing works, and the only difference is the shape of alpha:

SeasonalExponentialSmoothing            alpha shape (12,)   predict(level) raises   simulate raises
SeasonalExponentialSmoothingOptimized   alpha shape (12,)   predict(level) raises   simulate raises
SimpleExponentialSmoothing              alpha shape ()      predict(level) OK       simulate OK

For predict, the interval needs the alpha of the season each step falls in. k = ((steps - 1) // m) + 1 on the line above already gives the cycle number, so the season index is (steps - 1) % m.

For simulate, s_idx = i % m is already computed on the previous line for levels, and alpha needs the same index.

Scope:

• The scalar models are untouched. Their alpha is 0-d and their equivalent lines read differently (levels = alpha * ..., no [:, s_idx]), so they cannot be caught by this change — verified their predict(level=...) and simulate output is identical before and after.
• Point forecasts are identical before and after for all four models. Only the two paths that raised now return values.
• tests/test_models.py is 204 passed both before and after this change, since nothing that previously worked is touched.

Added a parametrised test covering both variants: it checks predict(h, level=[80]) returns ordered bounds and that simulate returns the expected shape. It fails on main for both and passes here.

Worth noting these paths appear never to have worked — the seasonal predict(level=...) and simulate have raised since they were added.

[codspeed]

Merging this PR will not alter performance

✅ 38 untouched benchmarks

---

_{Comparing VenishPaneliya:seasonal-es-per-season-alpha (43f427b) with main (8b28576)}