Do some profiling to find overhead for small arrays

[magnus-nomono]

When running benchmarks of a lot of small arrays I see that there is almost no difference in execution time between a 100_000 9 element arrays and a 100_000 23 element arrays. This indicates that we have some constant overhead dominating for small arrays.

I have already tried removing the python loop structure (notorious for being slow) from the measurements without it making much of a difference. Could it be overhead with cffi? Or something on the python side in the init?