Surrogate modeling: GP/RF interpolation from results table

[jc-macdonald]

Problem

After running a grid, users may want to interpolate between evaluated points to predict observables at untested configurations without running the expensive simulator.

Proposed

• fit_surrogate(results: ResultsTable, method="gp") → SurrogateModel
• Methods: Gaussian process (scikit-learn), random forest
• SurrogateModel.predict(config) → dict[str, float]
• SurrogateModel.uncertainty(config) → dict[str, float] (GP only)
• Could feed into run_adaptive as a cheap evaluator for multi-fidelity (Multi-fidelity support: cheap surrogate then expensive model #78)

Notes

Low priority. Ties into multi-fidelity support. Likely lives in an optional extra (trade-study[surrogate]).