Features:
* one small fix that we needed to patch for (atomicity of CallInstruction)
* removes the ambiguous scheduling warning

size is an expression so the result would be just some other expression.

We need more vector buffers for the tail loop!

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...

... really elegant.

Need to use handler from pymbolic as loopy codegen handlers
are too involved for standalone usage.

until poisson_neumann passes

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.