FVM: Choose better stencil for derivative in flux for D3Q27

As reported by @Tischler, the FVM discretization does not use the correct stencils for fluxes with derivatives in D3Q27. The result is not wrong, but uses more neighbors than necessary.

This merge request adds a test case and improves the stencil-choosing heuristic. It now produces the same two-point finite differences that a human would choose by optimizing the free weights later in the process.