SYCL: Don't use shuffles for top-level reductions

[masterleinad]

Partly reverts #6750. It turns out that using shuffles using more generic types such as in

Kokkos::parallel_reduce("kk_wup", N, KOKKOS_LAMBDA(int i, double& lsum1, double& lsum2) {
                      lsum1 += x_data_[i]*y_data_[i];
                      lsum2 += x_data_[i]*y_data_[i];},
                      result1, result2);

is much slower. Given that #6750 brought shuffle reductions only on par with local memory reductions, it makes sense to always use the latter for now. I have a slight preference for keeping the code for shuffle reductions (like we do for Cuda and HIP).

[masterleinad]

The SYCL build passed.

[ndellingwood]

@masterleinad does this need a changelog entry for 4.4? Looks like it may have impacts on performance?