[Tree unification 4/9] Honesty (held-out leaf re-estimation) on the kernel Tree

[jeongyoonlee]

Reimplement the honesty option (held-out leaf re-estimation) on the kernel Tree.

Depends on Issue 1 (scaffold). Issue 3b's IDDP path depends on this landing.

Scope

• Split the training data into a build split and an estimation split (estimation_sample_size, default 0.5).
• After the tree is grown on the build split, do a post-build pass re-estimating each leaf's per-group P(Y=1) on the estimation split — replacing the legacy honestApproach / fillTree — using Tree.apply to route estimation rows to leaves and rewrite each node's value.
• Respect honesty=False (default) as a no-op.

Verify

• Honesty parity vs UpliftTreeClassifier(honesty=True, ...) on a fixed seed.
• IDDP path (which forces honesty=True) works end-to-end.

Reusable templates

• Legacy: uplift.pyx honestApproach / fillTree (~615–883).
• Leaf routing: kernel Tree.apply.

Part of #945.