Spatial graph partitioning

[TDFSE]

Enhancement based on #4

• Average improvement of 33% compared to baseline on same system
• Partitioning disabled for <500,000 pixels - overhead caused performance reduction rather than enhancement
• Partition size of 32x32 is fixed - tried adaptive and aspect-ratio-aware adaptive scaling from 8x8->64x64, but 32x32 gave best results across the board
• Using outside datasets, have seen improvements with finding largest interior rectangle when compared to results from https://github.com/OpenStitching/lir

A/B comparison results using an NVIDIA RTX 4090:
comparison.txt
ab_comparison.json
performance_summary.json