JointStiffnessDriver should use InverseDynamics in gravity comp mode

[RussTedrake]

The current implementation turns off gravity on the specified bodies. This seems elegant, and is correct for many cases. But this has one problem: the effort limits on the input port are now incorrect.

To correctly model the input limits, we need to use the InverseDynamics system with InverseDynamicsMode == kGravityCompensation to compute the gravity comp forces, and pass those in to the input port. (And we should leave gravity enabled on the plant).