Add the possibility to warm up before measuring the throughput

[fwyzard]

This is implemented for most back-ends:

• alpaka
• alpakatest
• cuda
• cudacompat
• cudadev
• cudatest
• cudauvm
• fwtest
• hip
• hiptest
• serial
• sycl
• sycltest

The missing ones are those that I don't have any experience with:

• kokkos
• kokkostest
• openmp
• stdpar

[fwyzard]

This is an alternative version of #398 that targets the sycl backend.

It seems cleaner than #398, or at least less intrusive.

[fwyzard]

@makortel I'd like to merge this and use it as the baseline for benchmarking the ongoing Alpaka and SYCL developments. Any objections, comments, etc. ?

[makortel]

Visually looks ok. The remaining programs can be dealt with later.

[fwyzard]

Rebased, and fixed the conflicts.