feat: FittedImputer string_sentinels field, transform wiring, and serialization round-trip

[DEVunderdog]

Parent

#176

What to build

Add string_sentinels: dict[str, list[str]] as a dataclass field on FittedImputer (default empty dict) and wire it through transform and serialization:

• transform(): pass string_sentinels=self.string_sentinels to the existing _resolve_effective_nulls call, so replace-semantics string sentinel normalization happens at transform time on new data — including test-split data in production.
• to_dict(): serialize string_sentinels as a plain JSON-compatible dict.
• from_dict(): restore string_sentinels; a payload without the key defaults to {} for backwards compatibility with imputers serialized before this feature.

Acceptance criteria

• FittedImputer.string_sentinels field exists, defaults to {}
• transform() calls _resolve_effective_nulls(df, string_sentinels=self.string_sentinels)
• to_dict() output includes the "string_sentinels" key
• from_dict() restores the field; a dict without the key deserializes to {}
• A FittedImputer constructed with string_sentinels={"status": ["N/A", "missing"]} and serialized via to_dict() / from_dict() produces the same transform() output as the original object when the test DataFrame contains "N/A" in status
• A FittedImputer deserialized from an old-format dict (no string_sentinels key) transforms data identically to one constructed with string_sentinels={}
• No change in behaviour for FittedImputer instances where string_sentinels is empty
• Unit tests in tests/unit/imputation/test_fitted_imputer.py covering all criteria above
• Docstrings updated per ADR-0034

Blocked by

• Blocked by feat: extend _resolve_effective_nulls with user-declared string sentinel replace semantics #178