Host to device copies should be blocking

See https://github.com/JuliaGPU/OpenCL.jl/pull/243

From a quick glance, I didn't see where we're synchronizing (even for GPU -> CPU copies, which should be synchronizing already), so we should double-check here.

Start by reviewing the referenced OpenCL.jl pull request 243 and tracing how oneAPI host-to-device and GPU-to-CPU copies are synchronized. Done means confirming the synchronization behavior and ensuring host-to-device copies are blocking, with GPU-to-CPU copies remaining synchronized.

Written by the indexing model from the issue text.