feat(mice): add binary logistic FCS kernel

[alisadeghiaghili]

Summary

• Add a default MICE conditional kernel for categorical targets with exactly two observed levels: sklearn logistic regression plus chain-seeded Bernoulli draws.
• Preserve binary pandas Categorical, nullable boolean, and nullable string target dtypes; preserve observed cells and caller inputs.
• Route default binary targets through the manual FCS path even without typed-result/history options, while retaining existing numeric Bayesian-ridge and custom-estimator paths.
• Publish per-target typed-plan metadata: binary_logistic_fcs; provenance distinguishes numeric, binary, and multiclass/ordinal contracts. Multiclass and ordered categorical targets with more than two observed levels remain explicitly limited ordinal-encoded Bayesian-ridge approximations.
• Preserve where structural missingness, predictor-matrix selection, numeric bounds, and typed MI schema semantics.

TDD and verification

• Red public contracts failed under the prior implementation because every categorical target was labeled and processed as ordinal-encoded Bayesian ridge.
• New public tests cover string, nullable boolean, and ordered Categorical binary targets; support-only Bernoulli outputs; deterministic repeated seeded chains; dtype/observed-value/caller preservation; typed plan/provenance/warnings; where, predictor matrix, and numeric bounds interaction.
• Relevant MICE/MI/oracle tests: 172 passed.
• Ruff clean; mypy missingly/impute.py 0 errors; installed-module impute doctest 93 attempted/0 failed; diff-check and added-lines <=88 clean.

Statistical scope and risk

This is a bounded #71 capability, not a claim of full FCS validation, R mice parity, or #79 inference support. The frozen R mice oracle is numeric-only and is not used as categorical parity evidence. Logistic estimation falls back only on recoverable backend errors to an observed Bernoulli-rate draw with an actionable warning. Default binary completed values may differ from the prior ordinal-encoded Bayesian-ridge approximation; this is the intentional documented behavior correction. No coverage threshold, omission, exclusion, or pragma changes.

Baseline official main artifact: 92.50% (3846/4158), 312 missed. CI artifact determines any delta.

Refs #71
Refs #69
Refs #60