Speed up qbip

[jgostick]

This PR substantially improves QBIP and trapping performance while preserving existing results.

• Simplifies QBIP heap entries to contain only pressure and a flat voxel index.
• Eliminates per-voxel neighbor-list allocations.
• Separates invasion traversal from sphere drawing.
• Skips spheres already contained within the invaded region.
• Accelerates sphere rasterization using distance lookups and scan-line traversal.
• Avoids allocating disabled pressure and size outputs.
• Uses compact sequence and intermediate-array dtypes.
• Optimizes trapping with flat heap indices, direct neighbor insertion, smaller equal-sequence batches, fewer full-image temporaries, and direct mask generation.

Representative benchmarks showed approximately 10× faster QBIP and 12.5× faster trapping. A 500³ QBIP simulation completed in about 8.4 minutes, while the optimized trapping pass on a 220³ image fell from 13.5 s to 1.08 s. Memory use is also substantially reduced, particularly when optional outputs are disabled.