Fix loading of YOLOv8 sparse models on CPUs

[anmarques]

Small change to load models initially into CPUs. The model is transferred to GPU if needed later. Without this change the model failed to load on CPU-only systems.

Test plan:

1. I made sure that I could load the model in CPU-only machines and export to onnx, which motivated this fix.
2. In a GPU system I evaluated a dense model (zoo:cv/detection/yolov8-s/pytorch/ultralytics/coco/base-none) and made sure GPUs were used as expected
3. In a GPU system I evaluated a sparse model (zoo:cv/detection/yolov8-s/pytorch/ultralytics/coco/pruned50_quant-none) and made sure GPUs were used as expected
4. In a GPU system I ran a training command for a sparse model (zoo:cv/detection/yolov8-s/pytorch/ultralytics/coco/pruned50_quant-none) and made sure GPUs were used as expected
5. In a GPU system I exported a sparse model to onnx (zoo:cv/detection/yolov8-s/pytorch/ultralytics/coco/pruned50_quant-none)