Optimize combineShapesContinuousContinuous

[skosch]

The continuous/continuous case of algebraic combination (aka "convolution") is currently really slow. There are several opportunities for quality and speedup. At a minimum:

• Intelligent downsampling before starting the convolution
• Better target points (via some sort of clustering or binning?)
• Use of triangular kernels for adding up the results, and a tree structure (B-tree?) for the target nodes. The current implementation uses Gaussian kernels and is O(M × N) sloooow.

[OAGr]

I think I feel fairly uncomfortable about convolution and the challenges it seems to bring. It seems like it introduces a fair number of edge cases, and it's not obviously it's more worthy of investment than sampling at these points.