Introduce consistent mechanics for handling NAN

[llaniewski]

The current mechanics of Failcheck is cumbersome and not easy to extend. There are several places in the code (eg. Globals, DEM forces, different outputs) where code can cheaply check for NAN's, but here is no way to handle them.

We could introduce an element:

<OnNAN>
<VTK Iterations="1"/>
<Solver Iterations="10"/>
</OnNAN>

Which would register handlers, that the rest of the code can use in case of emergencies.

That could describe what you want to do when we see NAN anywhere. It could even have stuff like ignoreGlobals="true" to have more control. Failcheck would be still there, but the only function of it would be to explicitly check for NAN's in the domain.

FYI: @TravisMitchell