fix(transformer): align sklearn feature-name contract

[alisadeghiaghili]

Summary

• Makes MissinglyImputer.get_feature_names_out() conform to the sklearn transformer contract by returning a one-dimensional object numpy.ndarray in fit-time order.
• Adds public regression coverage for feature-name dtype/values and sklearn.clone, verifying cloned estimators retain constructor parameters but no fitted state.
• Replaces nullable fitted-state assumptions with typed, explicit invariants for strategy-specific sklearn imputers, PMM training data, and tree model dictionaries; targeted transformer mypy is now zero-error without ignores or suppressions.
• Repairs the stale strategy-validation doctest with a deterministic semantic assertion.

Coverage integrity

Official main baseline: 92.61% (3798/4101), with 303 uncovered lines. This PR changes no coverage threshold, omission, exclusion, pragma, or denominator configuration. The Python 3.12 CI artifact will determine the exact delta.

Verification

• Passed: 51 transformer tests.
• Passed: full Ruff, zero-error targeted mypy for transformer.py, installed-module transformer doctests (26 passed), diff-check, and changed-lines <= 88 columns.

Risk

get_feature_names_out() changes from a Python list to sklearn's standard ndarray output. Values and fit-time order remain unchanged. Other changes make impossible fitted-state corruption fail explicitly instead of leaking nullable/object internals.

Refs #78
Refs #60
Refs #69