GPU-compatible reduction of speeds in step size computation

[benegee]

Can we write this as a mapreduce? E.g. something like:

AcceleratedKernels.mapreduce(max, 1:num_elements, backend) do element
    max_scaled_speed_element(u, meshT, equations, dg,
                                                          contravariant_vectors,
                                                          inverse_jacobian,
                                                          element)
end

AcceleratedKernels does use Polyester for it's mapreduce on the CPU IIRC (grrml) so we could test the performance against @batch, another alternative is tmapreduce from OhMyThreads.

Originally posted by @vchuravy in #2590 (comment)

---

Needs to use backend through and maybe we can just use @invoke instead of this manual handling?

Originally posted by @vchuravy in #2590 (comment)