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.

I've already mentioned this problem in [!314] but I think more
visibility for this problem is useful

The SumfactKernel object stores a tuple of transformations that will be
registered in the realization and applied in the localoperator. These
transformations can change the name of the sum factorization kernel
function. This way it should be possible to use autotuning for picking
transformations.