test: end-to-end and FittedImputer tests for adaptive MICE

[DEVunderdog]

Parent

#91

What to build

Add integration and serialisation tests that verify the fully adaptive MICE path end-to-end, after all four behaviour slices (#158, #159, #160, #161) are merged.

test_imputation_end_to_end.py — one new case:
Dataset with multiple MAR-suspect columns where at least one column has a non-linear relationship with the others. Assert:

• Final imputed output contains no nulls
• Every MICE column's ColumnImputationRecord.signals contains an estimator entry
• Every MICE column's ColumnImputationRecord.signals contains a convergence-status entry (either the warning or confirmation of convergence)

test_fitted_imputer.py — two new cases:

• transform() on a MICE-backed FittedImputer produces no nulls in the output (covering the new estimator variants — RandomForestRegressor, GradientBoostingRegressor)
• to_dict() / from_dict() round-trip preserves the fitted IterativeImputer correctly for each new estimator variant (same contract as the existing MICE serialisation test, extended)

Acceptance criteria

• End-to-end test with a non-linear MAR-suspect dataset passes: no nulls in output
• All MICE column records in that test have a non-empty estimator signal
• All MICE column records in that test have a convergence-status signal
• transform() no-null test passes for RandomForestRegressor-backed MICE
• transform() no-null test passes for GradientBoostingRegressor-backed MICE
• to_dict() / from_dict() round-trip test passes for both estimator variants

Blocked by

• Blocked by feat: MICE estimator selection via NonlinearityTag + Unpredictable all-fallback #158
• Blocked by feat: dynamic max_iter and tol for MICE via seven-signal framework #159
• Blocked by feat: skew-driven initial_strategy for MICE IterativeImputer #160
• Blocked by feat: n_nearest_features selection for large MICE blocks via Pearson correlation #161