Add performance tests for an overhead of a kernel call

[fjarri]

With or without the preparation step. The latter is especially important for CUDA, where we may calculate the local size during preparation.