Implement Stacking module wrapping ArviZ.jl + Optim.jl

[jc-macdonald]

Summary

Port stacking.py for Bayesian model stacking and score-weighted ensembling.

Functions to Implement

Function	Description	Julia deps
stack_bayesian(loo_dict)	LOO/WAIC stacking weights	ArviZ.jl compare()
stack_scores(scores, directions)	Simplex-constrained score optimization	Optim.jl (constrained LBFGS or IPNewton)
ensemble_predict(predictions, weights)	Weighted average of model predictions	LinearAlgebra (stdlib)

Design Notes

• ArviZ.jl is a meta-package providing InferenceObjects.jl, MCMCDiagnosticTools.jl, PosteriorStats.jl.
• stack_scores needs a simplex constraint (weights ≥ 0, sum = 1). Optim.jl supports box constraints; project onto simplex after each step, or use a reparameterization.
• Package extensions: ArviZ.jl dep only loaded when user has it.

Acceptance Criteria

• All three functions exported and documented
• Weights sum to 1, all non-negative
• Round-trip test: stacking uniform models gives uniform weights