Investigate faster collision mesh generation/loading options

[kring]

As mentioned in #509, computing physics meshes for collision is a significant chunk of our tile load time, so we should avoid it when we can. When we do need it, however, it would be nice if it were faster. So this issue is to do some research into how this can be achieved, including but not limited to:

• Use Chaos instead of PhysX... is it faster?
• PhysX has a flag, EPhysXMeshCookFlags::FastCook, which it says can be used to "Prioritize cooking speed over runtime speed". How do we get Unreal Engine to invoke PhysX with that flag and what are the implications?
• Can we save time - and still get a useful result - by creating convex collision meshes instead of the fully-detailed ones?
• Similar to the above but even more extreme, can we use the tile bounding boxes as coarse collision volumes and will that be good enough for some use-cases?
• We probably don't want to put Engine-specific collision formats in 3D Tiles, even as an extension. But what extension could we add to 3D Tiles that would allow us to accelerate client-side collision mesh generation?
• Probably lots more.