[WIP] Neural Geometric Level of Detail demo

[duvenaud]

This is an attempted re-implementation (with @cyntsh and @chenzizhao , and with advice from @tovacinni) of NGLOD, which in principle should give us a real-time raytracer for implicit shapes.

This should be a good demo for Dex, because one of the main technical contributions of the paper is essentially a GPU-friendly implementation of the List monoid, although they frame it as using a parallel exclusive sum to compute the indices for flattened concatenated lists. In principle the Dex compiler should be able to automatically generate parallel code for this, although it might be hard to match their CUDA code.

Right now it is way too long, and repeats most of the ray tracer demo.