Fix OneHotEncoder metadata dropping generated indicators (#1106)

[avinashtrivedix]

Fixes #1106

Description

This PR resolves a metadata omission bug in OneHotEncoder where newly generated dummy columns were inadvertently dropped if their generated names collided with the original categorical input column names.

Changes Made

• Modified _add_new_feature_names() in feature_engine/encoding/one_hot.py.
• Reversed the order of operations in the list comprehension. The original categorical variables are now filtered out before the new dummy feature names are appended, entirely preventing the naming collision.
• Left the transform() method untouched, respecting the architectural updates already pending in the narwhals-migration branch.

Testing

• Verified the metadata fix locally against the minimal reproducible example provided in the original issue.
• Executed the existing test suite (pytest tests/test_encoding/test_onehot_encoder.py); all 29 tests pass successfully against the current main branch.

[avinashtrivedix]

I noticed the Python 3.9 pipeline is failing due to _SklearnCheckInputWrapper exceptions across unrelated modules (test_time_series, test_imputation, etc.). Since the style checks are green and the test_onehot_encoder.py suite passed successfully on the 3.10 and 3.11 runners, this appears to be a separate upstream environment issue on main. Let me know if you need any adjustments.