XPU: register native pinned host memory with the SYCL runtime
Native pinned buffers (DS_PIN_MEMORY_BACKEND=native) were mlock-only on XPU
because XPU_Accelerator inherited the no-op register_host_memory hooks, so
H2D/D2H stayed at pageable bandwidth. Implement the hooks via SYCL's
prepare_for_device_copy/release_from_device_copy and make the pin_memory
benchmark accelerator-agnostic so the gain can be measured on Intel GPUs.
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
