Add spatial and grouped validation with regional error analysis

[unes07]

Problem

A random row split does not demonstrate that recommendation models generalize to unseen sites or regions. Geographic autocorrelation and uneven crop coverage can make headline metrics overly optimistic.

Proposed work

• Group records by site and add commune-, province-, or region-held-out evaluation.
• Compare random-split and spatial-generalization results using identical features and targets.
• Report errors by crop, nutrient target, region, soil-data completeness, and zero-output frequency.
• Record split assignments so evaluations are reproducible.

Acceptance criteria

• No site appears in both training and evaluation groups for grouped tests.
• Spatial split definitions and coverage are published.
• Reports compare random and spatial results side by side.
• Sparse or unsupported regions are clearly identified.
• Documentation avoids broad reliability claims not supported by held-out geographic results.

Depends on #2 and #3. Uses cleaned and versioned data from open-turba/turba-data#5, open-turba/turba-data#6, and open-turba/turba-data#7.