Extend non-fitted error check to cover inverse_transform

[TemidayoA]

What this does

Part of #420 (the inverse_transform half — get_feature_names_out checks were added in #519).

The shared estimator check check_raises_non_fitted_error only verified that transform() raises NotFittedError before fit(). Every transformer that implements inverse_transform already guards it (via _check_transform_input_and_state / check_is_fitted), but no test covered that behaviour, so a regression would go unnoticed.

This extends the shared check to also call inverse_transform on an unfitted clone. Transformers that deliberately don't support inversion (OneHotEncoder, StringSimilarityEncoder) raise NotImplementedError, which the check accepts.

Because the check runs through check_feature_engine_estimator, all transformers across all modules are covered automatically — no per-transformer test changes needed.

Verification

• Full test suite: 2001 passed
• Sanity check that the extended check bites: a dummy transformer with an unguarded inverse_transform fails it; real transformers pass
• flake8 clean on the changed file

[TemidayoA]

Note on the CI failures on the first run, for whoever reviews:

• test_feature_engine_py39: all 51 failures are urllib.error.HTTPError: HTTP Error 403: Forbidden in the test_sklearn_wrapper dataset-download tests — a network flake on the runner, unrelated to this PR's diff (the non-fitted checks themselves passed; 1950 tests passed in the same job). Pushed an empty commit to retrigger.
• test_type: also failing on main's latest pipeline (build 28828), and this PR only touches tests/, which mypy doesn't check — pre-existing.
• upload_codecov: downstream of the failed job.

[solegalli]

Thanks for the PR @TemidayoA

Failing on both NotFittedError and NotImplementedError is a bit dirty, because we won't know if it fails for one or the other, so it won't detect changes that we do looking forward. If there are transformers that have the attribute inverse_transform and it's not implemented, we need to fail specifically for those with the NotImplementedError and for the rest with the NotFittedError

[TemidayoA]

Thanks @solegalli, that's a fair point — the tuple would have hidden exactly the regressions this check exists to catch. Fixed in the latest commit.

The check now determines, per transformer, which error it should raise and asserts only that one:

• inverse_transform is treated as not implemented when a fitted clone raises NotImplementedError (OneHotEncoder, RareLabelEncoder, StringSimilarityEncoder) → asserts NotImplementedError.
• every other transformer → asserts NotFittedError.

I detect implementation status on a fitted clone rather than a hardcoded list, so it stays correct for the conditional cases too (e.g. SklearnTransformerWrapper, which checks fit first and so raises NotFittedError before fit regardless of the wrapped transformer, and MeanEncoder whose default config is implemented).

I confirmed it's now strict: a transformer whose inverse_transform is implemented but drops its fit-guard (raising NotImplementedError before fit) now fails the check, where the tuple passed it. Full suite: 2001 passed, flake8 clean.

The classification fits on test_df(categorical=True, datetime=True) so every transformer that reaches this check can be fitted — the same input the neighbouring get_feature_names_out check already uses.