fix(optimization): return inf when sampler picks no-feature config

[SAY-5]

Closes #637.

mlforecast_objective's objective runs config_fn(trial) and hands the resulting MLForecast config straight to the model fit even when the sampled hyperparameters specify no input features (lags, lag_transforms, and date_features all None). That produces a zero-column feature matrix and sklearn raises:

ValueError: at least one array or dtype is required

The v1.0.2 objective guarded against this case (optimization.py:78-83) and returned np.inf so Optuna treated the trial as worst-case and moved on. The guard was dropped in 1.0.3 (commit 7526c56) without a follow-up. This restores it.

Patch

Matches the snippet the reporter validated locally, plus a trial.set_user_attr("config", ...) so the skipped configs still show up in study.trials_dataframe() for diagnostics, keeps the existing diagnostics behavior consistent with successful trials.

[CLAassistant]

All committers have signed the CLA.

[codspeed-hq]

Merging this PR will not alter performance

✅ 12 untouched benchmarks

---

_{Comparing SAY-5:fix/optimization-no-features-637 (58c74d2) with main (bed599b)}

[SAY-5]

All 14 other matrix runs (macOS / Ubuntu / Windows × Python 3.10, 3.13 + macOS/Windows 3.14) pass. Only all-tests (ubuntu-latest, 3.14) hit the 6h0m runner cap during dependency install/test (Python 3.14 is freshly released and the wheel resolver is slow on Ubuntu); looks like an infra timeout rather than a real failure on this small one-config-skip patch. A retest should clear it.