update handling of self=FALSE in getParents and use in setupMargNodes

[paciorek]

This combines the tentative work in PR #1630 and PR #1631 , plus removes what should be redundant calls to setdiff in setupMargNodes, now that self=FALSE does omit all nodes that are provided as the input nodes to getParents.

We are considering including this PR in version 1.4.3.

We are thinking that to do so we probably want to run all the extended nimDerivs testing not included in CI.

[paciorek]

The test-ADforwardSolve failure seems to be a heisenbug. It is fine when I run it locally.

[perrydv]

This looks like a tolerance issue, what has been a pain for AD testing and seemingly can arise for inscrutable reasons like R or library or OS or compiler versions. I'd suggest running tests on devel to determine if this failure arises from any changes to the CI testing setup.

[paciorek]

Ok, I've run nimbleQuad tests and the extra (non-CI-enabled) AD-related tests. Things look fine (there are a small number of isolated test failures but they seem to be finicky/heisenbugs and also seem to occur on current nimble).

So I think I will proceed with including this in the release tomorrow.

[paciorek]

Ok, moving ahead with this for 1.4.3.

This PR also updates some confusing help text language for setupMargNodes and a change to the code of setupMargNodes() (see discussion in PR #1630) and reworks use of setdiff in setupMargNodes now that we have changed behavior of self=FALSE for getParents (see PR #1631).