Speed up lazy vector-backed LP and NLP solves

[daggbt]

Summary

Preserve lazy VectorVariable metadata further into LP and SciPy solve setup so single-vector problems avoid unnecessary scalar materialization and compile overhead.

What changed

• propagate lazy vector metadata through LP extraction, bounds handling, and solver setup
• add a contiguous single-vector compile/cache path for SciPy NLP solves to avoid repeated per-element name generation and index lookups
• keep solution values lazy on the solver result path while preserving warm-start behavior
• add regression coverage for the lazy-cache/compiler paths and refresh benchmark artifacts/docs

Why

These changes carry the constructor-side lazy vector win through problem.solve(), reducing cold-start overhead for large single-vector NLPs while keeping the Python implementation path intact.