SYCL: Use host-pinned memory to copy reduction/scan result

[masterleinad]

For dot-type kernels, I see an improvement between 10% and 50% in runtime when using host-pinned memory with memcpy vs. device memory with sycl::queue::memcpy on Intel GPUs. This is analogous to what we are doing for Cuda and HIP.

Noticed while looking at #5334.

[masterleinad]

CUDA-12.2-NVHPC and OPENACC-NVHPC-CUDA-12.2 were timing out, OPENMPTARGET-Clang failed openmptarget.partitioning_by_vector which we see regularly. Everything else, In particular the SYCL build, passes.

[masterleinad]

Retest this please.

[masterleinad]

All CI is passing.