The old formula contained some 'dim's instead of '2's,
which obviously works for 2D....

Tricky... some grad components depended on the sumfac kernels of the
other directions instead of their input...

It is not (yet!) suitable for our needs.

This makes the order of 'gradvec' and 'quadvec' consistent with the
other tests.

Note: sumfact_poisson_fastdg_gradvec_nonquadvec does not work!

At the moment loopy tries to acces a vector class variable and throws
a not implemented error. Probably needs some changes in the horizontal
add code.

For some reason, the patch refuses to schedule the kernel for sumfact_poisson_dg_symdiff.