Address quadratic inverse handling - #9371

I don't fully understand the internal data structures and how inverses are generated but here are my observations using this code.

• My export is noticeably faster, on par with the v0.8.0 implementation, but I don't have hard numbers to compare
• I get better performance on my C++ IFC export implementation if I bundle up all the RelatedObjects with their RelatingObject (or whatever the attributes are called) in a map, and then create the IfcRel directly one time, instead of calling file.addRelatedObject<>.
• On the import side, using Blender/Bonsai, it is glacially slow using the v0.9.0 implementation at the fix from this branch. It was just as slow without this change also. The example bridge model I shared loads in less than a minute when backed by v0.8 and after 30 minutes it's still loading when backed by v0.9