perf: accelerate model build workflow

[koen-vg]

Summary

• reuse exact region/resource-class cell mappings across spatial preprocessing
• vectorize LUC, land-area, water, MIRCA, multi-cropping, and crop-link operations
• replace repeated LUIcube grassland polygon traversals with mapped reductions

Impact

On a clean default build through build_model with -j8, wall time fell from 150.94 seconds to 90.91 seconds and peak reported RSS fell from 5.58 GB to 4.70 GB. Downloads were excluded from both runs. The subsequent LUIcube optimization reduced that rule from 11.41 seconds and 2.37 GB RSS to 0.63 seconds and 711 MB, with byte-identical output; this improvement is not yet included in the full-run figure.

Validation

• full 261-job workflow through build_model
• strict equality of the 305,411-link PyPSA network
• byte-identical LUC, land-area, MIRCA, multi-cropping, and water outputs
• xarray-identical resource-class output
• 89 focused tests and the integration dry-run
• Ruff, Snakefmt, REUSE, and pre-push hooks