Candidate fix to #1564 (but in getConditionallyIndependentSets)

[perrydv]

I am throwing this up as a candidate fix. It fixes Issue #1564 and it passes test-setupMargNodes.R.

We should probably tests in (at least) nimbleQuad (in development at this time) and nimbleSMC,

If this seems to work, I will document more about the change and add a regression test.

The basic idea is: don't mark deteministic nodes as touched during getConditionallyIndependent sets traversals of the graph. The reason is that as we are iterating over a set of parent (or presumably in some cases child) nodes, the recursions could pass through a deterministic node that then needs to be passed through by a later parent (or child) node of the original iterations. But if it has been "touched" (tagged as done), then it blocks activity. But in the case of a deterministic node, this is not really correct. So it takes a particular setup of nodes to trigger the issue. I think it is actually correct to never tag deterministic nodes, since the algorithm is fundamentally about stochastic nodes.

[paciorek]

For discussion - it's not clear to me why tests would go into nimbleQuad given this is really about cond. indep. sets.

I've added a (the?) regression test for the example in #1564.

I'll note that setupMargNodes is mostly about Laplace-related stuff but is used in MCEM too. So my plan was to keep it in nimble and not move to nimbleQuad.

[paciorek]

This passed testing previously; I just canceled the tests to avoid a bunch of test computation when adding the simple regression test.

[paciorek]

This passes the (limited new) tests in nimbleQuad and the nimbleSMC tests. I'm going to merge in.

(Actually a test-quadrature.R test is failing, but that was apparently failing before.)