[FEA]: Add more benchmarks for thrust::transform

[bernhardmgruber]

We currently use the BabelStream kernels as memory bound workloads and a fibonacci kernel (per thread: read random index [0;42], compute fibonacci number, store result). The latter kernel is compute and shows high thread divergence.

In order to better assess regressions, we should add more benchmarks covering:

• compute bound, but convergent threads
• overwriting the input sequence 1:1
• reading from overlapping regions (e.g. something like adjacent difference, or a stencil operation)

[bernhardmgruber]

I think we should actually make them CUB benchmarks, since they should be included in continuous CUB benchmarking and tuning.

[bernhardmgruber]

We can probably postpone this, because thrust::transform seems to be performing well.