Share tree kernel: hoist group criterion base + unify builders (#945)

[jeongyoonlee]

What

Further modularize the tree kernel so the causal and uplift trees share more code, per the direction in #945. Two focused, behavior-preserving refactors:

1. Hoist the shared treatment-group criterion base into the _tree kernel.
NodeState / NodeSplitState / CausalRegressionCriterion — the per-group (control-vs-treatment) sufficient-statistics base that both families' criteria extend — moves out of causal/_criterion into a new shared module _tree/_group_criterion. The uplift criterion now depends on the shared _tree kernel instead of reaching sideways into the sibling causal package.

2. Unify the two forked builder skeletons.
causal/_builder and _uplift/_builder were each a ~570-line fork of scikit-learn's DepthFirstTreeBuilder / BestFirstTreeBuilder (~76% identical). The common skeleton now lives once in _tree/_group_builder as GroupDepthFirstTreeBuilder / GroupBestFirstTreeBuilder, with the family differences expressed as four overridable cdef hooks (all no-ops by default):

hook	causal	uplift
_node_is_leaf_extra	per-group min-size stopping	—
_before_node_split / _after_node_value / _before_build	—	parent-summary shrinkage threading

The family builders collapse to thin subclasses (~90 lines each) overriding only their hooks. The two now-empty _builder.pxd files are removed (the subclasses are self-contained).

Why

The criterion base was already shared by inheritance (uplift's criterion subclasses the causal one), but it physically lived under causal/, so uplift depended on a sibling rather than on the kernel. And the builder skeleton was copied verbatim in both families. This centralizes both in _tree/, the single home of the shared tree machinery — the layering the #945 epic was driving toward.

Behavior preservation

Both changes are behavior-preserving. A parity harness fits 37 trees — causal (causal_mse/standard_mse/t_test × depth-/best-first × group-size stopping) and uplift (KL/ED/Chi/CTS/DDP/IT/CIT/IDDP × plain / regularized / normalized / honest) — and asserts the fitted tree_ arrays (feature, threshold, value, children, impurity, counts) are byte-identical before and after each tier. Result: 37/37 identical after Tier 1 and after Tier 2.

Note: only the depth-first uplift builder is reachable from the public API (_KernelUpliftTreeClassifier fixes max_leaf_nodes=None); the best-first uplift subclass is kept for symmetry. The shared best-first retains sklearn's impurity <= EPSILON early-leaf — a no-op for uplift, whose node_impurity is the constant 1.0.

Verification

• Parity harness: 37/37 byte-identical trees (Tier 1 and Tier 2).
• tests/test_causal_trees.py, tests/test_uplift_trees.py, tests/test_uplift_trees_kernel.py (incl. the legacy-vs-kernel parity tests) — all pass.
• black clean.

Net effect

~1140 lines of forked builder collapse to one ~600-line shared skeleton plus two ~90-line subclasses; the criterion base is a pure relocation. _tree/ becomes the single source of the shared criterion base and builder traversal for both tree families.

Part of #945.