Get gradients working with uncollapsed maker SPs

[axch]

Currently, the derivative information for the output's influence on the latent variable, and both of those influences on the inputs, is simply lost.

Issues:

• Is it reasonable to take gradient steps on the made SP itself (if it has continuous latent state)?
• Gradient of what, exactly? The AAALKernels associated with stochastic makers take Gibbs steps, so the gradient of their weight is always 0.
  • Perhaps we need an explicit notion of logDensityOfData even for stochastic made SPs, in order to be able to take gradient steps with respect to it. (On a simplex, in some cases!)
  • On the other hand, what is the desired behavior if an uncollapsed AAA maker (that takes Gibbs steps) is a non-principal resampling node in a gradient-based proposal?

[axch]

The commentary on #455 provides a hint to this problem. When the uncollapsed kernel understands the local posterior well enough, it should be able to report its size (as a measure). This is what logDensityOfData computes.

We could then choose to move input (hyper-)parameters according to that feedback.