Oct Tree sampling and Autodiff about descartes.jl HOT 1 OPEN

The main question is how should Ck continuity be specified. We ostensibly want different functions to avoid call-site branches. The current notion of FRep can allow for propagation of an N parameter to use generated functions. e.g. Frep{2} gives a twice differentiable function using Rvachev clipping and FRep{0} will implement max/min variants for performance. That said I think the first component of this function parameterization is easy to implement on the GPU. The question is if we use ForwardDiff to accomplish these goals we will need to provide some sort of type overload on the GPU.

from descartes.jl.