Add Intel dGPUs and iGPUs capabilities to DEIMV2

[kprokofi]

This PR makes the core training/evaluation/inference pipeline device-agnostic so -d xpu works out of the box on XPU-enabled PyTorch builds.

Validated on an Intel ARC B70 device using torch 2.13.0+xpu / torchvision 0.28.0+xpu:

• XPU GradScaler scale/backward/step/update round-trip.
• train.py -d xpu --use-amp — 1-epoch training + COCO eval completed successfully.
• tools/inference/torch_inf.py — byte-identical output between CPU and XPU inference.
• torchrun --nproc_per_node=1 XPU DDP eval passed with the xccl backend.
• compileall, py_compile, and git diff --check all clean.