Native missing-value (NaN) support for kernel uplift/causal trees (prereq for #955)

[jeongyoonlee]

What

Prerequisite for the #955 switchover (tree-unification epic #945). Adds native missing-value (NaN) support to the kernel-backed uplift and causal trees/forests, mirroring scikit-learn's tree implementation closely. Today these estimators reject NaN in X; the switched-over public UpliftTreeClassifier / UpliftRandomForestClassifier need to accept it, so this lands the criterion-level work first, isolated from the compat/switchover mechanics.

How (follows scikit-learn's design)

The shared CausalRegressionCriterion (inherited by every uplift and causal criterion) now implements sklearn's per-feature missing-value machinery:

• NodeSplitState gains a per-group missing NodeState. init_missing(n_missing) accumulates the missing tail (sample_indices[end - n_missing:end], the sklearn contract) into both the flat sum_missing (used by StandardMSE) and the per-group state — still respecting the group-encoding NaN mask (isnan(y[i, k]) = "sample not in group k").
• reset / reverse_reset fold the missing block into the child chosen by missing_go_to_left; update scans only the non-missing range (end - n_missing); StandardMSE.children_impurity adds the missing squared-sum when routed left. All are no-ops when n_missing == 0, so clean-data behaviour is unchanged.
• The flat sum_missing allocation (init_sum_missing) is inherited from RegressionCriterion unchanged.

The Python tree layer follows sklearn 1.7:

• _support_missing_values + an allow_nan tag via __sklearn_tags__ (dense best splitter only — random/sparse don't carry the machinery). The vendored base's _get_tags()["allow_nan"] is dead under sklearn 1.7; the causal/uplift layers now use __sklearn_tags__().input_tags.allow_nan.
• fit validates with ensure_all_finite=False and threads _compute_missing_values_in_feature_mask(X) into builder.build(...).
• _prepare_data and both forests allow NaN through validation. inf is still rejected. The forest allow_nan tag is delegated automatically by sklearn's BaseForest.__sklearn_tags__.

Scope

• Enabled for: _KernelUpliftTreeClassifier, _KernelUpliftRandomForestClassifier, CausalTreeRegressor, CausalRandomForestRegressor (the criterion is shared, so causal benefits for free — as discussed for the epic).
• Not covered (consistent with sklearn): the random splitter and sparse input; and a feature whose non-missing values are all constant is still treated as constant (its missing-only split is not separately proposed).

Tests

New missing-value tests for both the uplift and causal trees/forests:

• acceptance (fit + predict with NaN, finite/valid outputs), the allow_nan tag + _support_missing_values (dense True / sparse False),
• all-NaN column and all-NaN prediction row,
• learned NaN routing — effect concentrated on the missing rows is recovered (NaN rows → high-effect leaf, observed rows → ~0),
• inf still rejected, forest fit/predict with NaN.

Full tests/test_uplift_trees_kernel.py + tests/test_causal_trees.py pass (154); black clean.

Part of #945. Prerequisite for #955.