Add kernel-vs-legacy uplift tree parity notebook (#945)

[jeongyoonlee]

What

Adds docs/examples/uplift_tree_kernel_parity.ipynb, a validation notebook for the tree-unification epic (#945) that demonstrates the experimental kernel-backed uplift trees reproduce the legacy implementation numerically.

The notebook shows, with executed outputs, that _KernelUpliftTreeClassifier / _KernelUpliftRandomForestClassifier match UpliftTreeClassifier / UpliftRandomForestClassifier bit-for-bit (max abs difference 0.0) on:

• Single-tree criteria — KL / ED / Chi (multi-treatment) and DDP / IT / CIT / IDDP (single treatment; IDDP forces the honest approach)
• Regularization + normalization and the honest approach
• Whole-forest predictions

plus a fit-time speed comparison (kernel ~3x faster on the sampled data).

The last section makes the parity conditions concrete: on genuinely continuous features the two diverge because they search different split grids, but discretizing the same features restores exact parity — confirming the difference is only the split-candidate grid, not the criterion math.

Notes

• Parity requires binary/low-cardinality features and legacy_max_depth = kernel_max_depth + 1 (both explained in the notebook).
• The kernel classes remain experimental and private (underscore-prefixed); they are imported here only to demonstrate parity. The public classes are switched over to the kernel later in the epic ([Tree unification 9/9] Switchover + delete legacy uplift.pyx + deprecations #955).
• Wires the notebook into the docs examples toctree (docs/examples.rst).
• CI does not execute notebooks; committed outputs are what render in the docs.