feat(ml): GPR as default model with uncertainty estimates and EI prop…

[ngholiza]

…osals

• Replace hardcoded RandomForestRegressor with GaussianProcessRegressor (default)
• GPR uses RBF + WhiteKernel kernel with Tikhonov regularisation (alpha=1e-6)
• LOOCV parity now returns uncertainty_std (σ) per prediction point
• Proposals use Expected Improvement (EI) acquisition with diversity constraint
  • exploitation: highest EI (best known region)
  • exploration: high σ (uncertain region, never visited)
  • diversity: spread across parameter space
• RF still used: (a) for Gini/MDI importance (b) when dataset > GPR_ROW_LIMIT=400
• All functions accept optional domain_config dict (features, targets, ml_model) enabling config-driven ML for multi-domain support
• Backward-compatible: etcher UI aliases (etch_rate, range_nm) preserved