In the /sumfact_poisson_dg_3d_unstructured_deg2_symdiff_nonquadvec_gradvec
test case this reduces the time spent in evaluation of it from >100s to 6s.

In many places we treat this as codegen-time information anyway.
For loopy, this is the necessary way to go.

Run the function call over multiple data sets.

Replace placeholders in template file with actual code. This makes it a lot
easier to restructure the benchmark.

This rewrite was necessary for the cases where the instruction contains a
reduction but the expression itself is not a reduction. Besides fixing this bug
it greatly reduces code duplication and increases code quality.

This just reverts to the old behaviour. The new implementation of loop
reordering doesn't need that anymore.

Still missing: Handling of reduction with haddsubst!

- Code cleanup
- Use loopys duplicate_inames
- 2D, 3D volumes and faces should all be working

This can be used to distinguish between kernels with the same name but
different transformations.