Optimization: Improve Manifold caching

We currently cache every node in the OpenSCAD CSG tree. However, when using the Manifold backend, this circumvents Manifold's ability to optimize evaluation of CSG trees.

Ideas include:

• Cache only certain node types (e.g. minkowski, hull)
• Estimate complexity an only cache nodes which are expected to be CPU intensive to evaluate

See some related discussion in #5485