Enable evaluation of treeNodes via sampling

[skosch]

Most of the time, convolution will be fast and give very nice results. But there are situations where sampling makes more sense. In particular, if the whole expression consists of a bunch of algebraic combinations, then it will be much faster to just estimate the overall minX/maxX, compile the whole tree into a JS function, and take a few thousand samples. And then there are the cases where it makes sense to sample a particular subtree (e.g. for pointwise products), but do the rest of the work on XYShapes directly. So for any given treeNode, we need to:

1. Have some kind of heuristic for what strategy to use to evaluate the subtree.
2. If we decide to evaluate via sampling, provide a way to do this:
   1. Easiest: compile the subtree into a Guesstimator string, let Guesstimator do the work. But this probably won't work, because we have functionality now that Guesstimator (probably?) doesn't have, like pointwise scaling etc.
   2. Better: compile the subtree into a JS function, and pass that to Guesstimator's sampler directly.
   3. Even better: compile the subtree into a JS function, and use a more sophisticated sampler. It might be relatively easy to lift code from WebPPL?

[OAGr]

We have a simple version of this now.