This is a rehash of #3637, hopefully more likely to be productionized.

Goals:

• Get as many lazy top-level unions as possible
• Push down transforms to reduce the # of transforms being done, mitigate precision concerns (esp. w/ the Manifold rendering backend introduced in Use Manifold for much faster & multithreaded CSG & minkowski operations #4533) + allow more unions to bubble up.

Note that I've improved Manifold's tree flattening in elalish/manifold#368. We could potentially ask Manifold for a lazy top-level union instead of its full rendering (it keeps a lazy CSG AST under the hood), which could be a better approach than this PR (it might still need to deal with the issue of transforms precision). Still, I wanted to get this draft PR out there as it works even w/o Manifold.