fix: preserve target scale in MFLES fallback

[Dresden (DresdenGman)]

Summary

Keep the original target scale in MFLES's existing short/constant-series fallback. The fallback currently runs after normalization/log transformation, overwrites the saved additive transform parameters, and returns transformed fitted values without applying the inverse transform.

For example, on main at 99f2408:

import numpy as np
from statsforecast.models import MFLES

model = MFLES().fit(np.full(10, 7.0))
print(model.predict(3)["mean"])                 # currently [0., 0., 0.]
print(model.predict_in_sample()["fitted"])    # currently ten zeros

model = MFLES().fit(np.array([5.0, 7.0]))
print(model.predict(3)["mean"])                # currently [1., 1., 1.]

The existing fallback is intended to repeat the last observation, so these cases should return 7 on the original target scale. With multiplicative=True, a short history such as [5, 7] also returns log-scale fitted values even though its forecasts are inverse-transformed.

Changes

• Preserve the mean and standard deviation saved before normalization, so predict applies the correct inverse transform.
• Inverse-transform the fallback's fitted values, as the non-fallback path already does.
• Keep the fallback trigger and ordinary decomposition/optimization paths unchanged.
• Add offline regression tests for additive/log-scale cases, positive/negative/zero/fractional constants, one-to-three-point histories, float32/float64, both MFLES/AutoMFLES public wrappers, and fitting a fallback series after a model with exogenous components.

This is independent of #911's forward/update design and #1215's multi-seasonal cycle fix. It does not add a forward method or change seasonal-cycle construction.

Validation

Python 3.12.12, NumPy 2.5.3, SciPy 1.18.1, pandas 2.3.3 and scikit-learn 1.9.1; StatsForecast built from 99f2408.

• New regression file against the earlier unmodified base d31d8ab: 39 failed, 7 passed. The affected upstream code is unchanged at 99f2408, where the minimal examples above still reproduce.
• After rebasing the patch onto 99f2408: 48 passed across the new regression file and the existing MFLES tests.
• Eight non-fallback controls (seasonal/nonseasonal × additive/log × with/without exogenous features): fitted values and forecasts exactly match unmodified main.
• Ruff check and whitespace validation pass.

python -m pytest -o addopts='' -q tests/test_mfles_naive.py tests/test_mfles.py
ruff check tests/test_mfles_naive.py python/statsforecast/mfles.py
git diff --check

The targeted command disables the full-suite coverage gate; this is not a full-suite coverage result. The existing demand-data-download integration tests were not run. Explicit log transforms of nonpositive data and seasonal decomposition output for degenerate series are outside this patch's scope.

[codspeed]

Merging this PR will not alter performance

✅ 38 untouched benchmarks

---

_{Comparing DresdenGman:fix/mfles-naive-original-scale (99fd875) with main (99f2408)}