feat(surrogate): GP/RF interpolation from results table (#82)

[jc-macdonald]

Closes #82.

Summary

Adds a new trade_study.surrogate module that fits scikit-learn surrogates (Gaussian process or random forest) to a ResultsTable and predicts any observable at unseen factor combinations.

API

from trade_study import fit_surrogate

model = fit_surrogate(results, factors, method="gp", seed=0)
model.predict({"alpha": 0.5, "beta": 0.3})         # {"y": ..., "z": ...}
model.predict_batch([{...}, {...}])                # {"y": ndarray, ...}
model.uncertainty({"alpha": 0.5, "beta": 0.3})     # GP only; RF raises

• Continuous factors are min-max scaled to [0, 1].
• Categorical / discrete factors are one-hot encoded against factor.levels.
• NaN rows are dropped per-observable, so partial sweeps still fit.
• method="gp" uses ConstantKernel * Matern(nu=1.5) + WhiteKernel with normalize_y=True; method="rf" uses RandomForestRegressor.

Packaging

• New optional extra: pip install "trade-study[surrogate]" (scikit-learn ≥ 1.3).
• Also rolled into the all aggregate.

Docs

• New docs/api/surrogate.md page, wired into the mkdocs nav between Stacking and Visualization, with an install hint for the optional extra.

Tests

13 new tests covering both backends, GP uncertainty, mixed-factor encoding, NaN row handling, and input validation. surrogate.py coverage: 100%. Total: 296 passing, 99.5% project coverage.

Follow-ups

• Unblocks Regime-conditional surrogate: interpolate factor recommendations across regime features #105 (regime-conditional surrogate) — that issue can wrap a per-regime SurrogateModel and a regime classifier on top of this infrastructure.