Fixes bug introduced by #3596 when handling NaNs.

[pepbos]

My Pr #3596 introduced a bug, or triggered a bug elsewhere, when handling NaN's in SmoothSegmentedFunction.

Not sure when NaNs will be injected, and if that even makes sense, but before #3596 that was OK, but currently main will raise an exception. This PR makes it fine again, as was before #3596.

I noticed when trying to visualize my result from a CMC, which raised an exception now.

---

This change is 

[nickbianco]

Do you think this might be hiding another issue? Is it possible that your CMC simulation/visualization is feeding this function a NaN value for x which then gets masked by returning zero here. Or do you think the changes introduced new NaNs somewhere?

[nickbianco]

I see, this was the previous behavior of the function. In that case, this is probably fine.

[pepbos]

Looks like the intention was for the smoothSegmentedFunction to return NaN for a NaN as input (which makes sense).
However it had three different cases:

• calcValue(NaN) returned NaN
• calcDerivative(NaN, 1) returned dydx1
• calcDerivative(NaN, 2) returned 0

Changed this PR to fix always returning NaN when the input is NaN, so:

• calcValue(NaN) return NaN
• calcDerivative(NaN, 1) return NaN
• calcDerivative(NaN, 2) return NaN

Note that without this PR an exception is raised if the input is NaN, which blows up some cases that were fine before.
I added the NaN case to the unittests.

[nickbianco]

LGTM. Thanks for the clarification @pepbos.

[nickbianco]

Also, you can ignore the Mac failures, that is an issue with the CI. Working on a fix for that.