Document honest estimation and pruning in the methodology

[jeongyoonlee]

Honest estimation became the default for CausalTreeRegressor and CausalRandomForestRegressor (#584/#1001) and UpliftTreeClassifier learned to prune inside fit (#1003/#1004), but both live only in the changelog and one notebook. A default that changes fitted models belongs on the Methodology page.

Adds two subsections under Tree-Based Algorithms:

• Honest estimation — what it is (structure grown on one half of the sample, leaf values re-estimated on the other, removing split-selection bias; Athey & Imbens 2016), which estimators offer it with which defaults (honesty=True on the causal trees, False on the uplift trees; estimation_sample_size=0.5), the bias/variance trade, and how forest honesty composes with the bootstrap rather than following Athey, Tibshirani & Wager 2019.
• Pruning — the uplift tree's validation-based prune() / prune_fraction, and the causal tree's cost-complexity ccp_alpha including "cv" (honest cross-validation over the cost-complexity path; requires honesty=True; prefer on a single tree).

Content restates the measured conclusions already in the changelog rather than re-deriving them. Docs-only; sphinx -W clean on html and latex.

First of the three P1 items from the docs roadmap.

🤖 Generated with Claude Code