There are now stencils related to elements and stencils related to vertices.
The element stencil for box e.g. contains all dof related to an element
whereas the vertex stencil contains all dofs related to a vertex (==dof).

Apparently returning sets by value is extremely slow. Return everything
by reference. Try use insert. Eliminate map in localjacobian.
* use stationary capability. We now have a speedup of 30%! im comparison
to stable at solver dominated 300x300 cells.

* Boundary volvars are only needed/updated once per fluxvars instance.
* The fluxes can be precomputed to be more efficient in composiitonial models.
* The density has to be outisde the gradient operator in Fick's law!

The primary variable switch now updated the global volvars. In the
newtonEndStep method in the model only those volvars need to be updated
now that switched primary variables.

* The immiscible folder contains files common to immiscible
  porousmediumflow. It now contains a general local residual for np flow

* The compositional folder contains files common to the compositional
  porousmediumflow models. It now contains the privarswitch base class.

All three diffusion coeffficients were previously computed on each
function call but only two are needed per phase.

The diffusion coefficient computation consists of three steps.
A scalar diffusion coefficient can be constant or computed depending
on primary and secondary variables in a constitutive law. This is
done for every volume variable once and thus in the volvars update.
In a second step an effective diffusion coefficient in computed
depending on saturation and porosity. In a third step (currently
not implemented) a fibre structure could lead to anisotropic
diffusion tensors. They could be provided in the current framework
as normalized tensors per SCV by the spatial params and then
scaled with the computed diffusion coefficient (see e.g. Linniger 2012).

This is used to update the secondary variables or perform
variable switches

The flux variables are now named according to the physical processes they represent.
There is an advection type (possbibly darcy, forchheimer, stokes) where currently darcy's
law is implemented for CC TPFA. There is a diffusion type. This implements fickean diffusion.
And a heat conduction type that could be implemented as Fourier's law. The latter two types
are commonly of elliptic type like (D grad c) or (lambda grad T) where the first one is only elliptic
for darcy's law (K grad P). The flux vars are specialized for models via the properties
EnableAdvection, EnableDiffusion, and EnableHeatConduction. The contrasts (K, D, lambda) can be
thus possibly precomputed.

* introduce a variable switch property
* derive the 3p3c switch from a base primary variable switch

Instead of variables we have classes computing fluxes
according to laws like darcy's law / fick's law / fourier's law.