Profile DelaunayNN interpolation caps

[Jammy2211]

Summary

• add a full JAX mapper-table benchmark for DelaunayNN
• record versioned CPU JSON and PNG results
• document the workspace cap audit and the cap-32 decision
• link the profile and evidence from the repository indexes

Results

At 1,200 vertices and 15,000 queries, median end-to-end CPU times were:

• Delaunay: 0.066746 s
• DelaunayNN cap 16: 0.192315 s
• cap 24: 0.239952 s
• cap 32: 0.317660 s
• cap 64: 0.562402 s

Cap 32 is the smallest tested safe choice in the 101-model workspace audit. Cap 64 adds substantial cost without an observed coverage benefit. An A100 run remains future profiling work.

Dependency

Profiles the core implementation in PyAutoLabs/PyAutoArray#441.

Validation

• ruff check .
• ruff format --check .
• generated README check
• smoke import
• git diff --check