[Box] Numeric differentiation not correct if volume variables depend on dofs other than the scv dof

Bug reported here by Dmitry Pavlov on the mailing list.

When deflecting the solution to compute the numerical derivative of the element residual in the box method, currently only the volume variables associated with the deflected dof are updated. However, since volume variables may depend on all dofs of the element, also the other volume variables on the element have to be updated.

This leads to an inexact Jacobian. Fixing the issue may lead to a significant overhead for most models as volume variable updates are typically the most expensive step in the local residual calculation.