Add DelaunayNN workspace coverage

[Jammy2211]

Summary

• add a workspace reconstruction comparison between Delaunay and DelaunayNN
• add a production-like cap audit over 101 ray-traced mass models
• include both scripts in smoke coverage

Findings

For 1,200 mesh vertices and 4,421 data-grid rows, observed maxima were 27 neighbors and 25 cavity triangles on the main grid, and 21 neighbors and 19 cavity triangles on split points. Cap 16 overflowed substantially and cap 24 still overflowed four rows in the worst model; cap 32 covered the complete audit.

DelaunayNN source reconstruction was numerically close to Delaunay: relative L2 error 9.884215e-05, correlation 0.998595442.

Dependency

Exercises the core implementation in PyAutoLabs/PyAutoArray#441.

Validation

• both scripts pass Ruff formatting and lint
• full default cap audit passes
• git diff --check passes