CUDA backend temporaries should respect CUDA alignment guidelines

[anstaf]

A minor problem with this: vertical advection is the only our regression that use temporaries in cuda; our perftests use power of two domain sizes. Hence we need yet another test to track the performance improvement of that change.

[havogt]

The idea is to align the first domain point of each CUDA block. Currently, we align to the first halo point of each block.

We came to the conclusion that this is likely only a minor performance improvement for only very specific stencils:

• pure vertical stencils are well aligned as they don't have a halo (no improvement)
• pure horizontal stencils should use the cuda_horizontal backend or ij caches (no improvement)
• for mixed stencils it depends probably on the stencil which alignment is better (maybe an improvement)