examples: use public COCO128 dataset and YOLOv8n model in YOLO tutorial

[ashishsoni-ai]

YOLO tutorial: use public COCO128 dataset and YOLOv8n model

What

Fixes examples/tutorial_image_detection_yolo.ipynb so it runs end-to-end without access to
private assets, following the same pattern as #613 (COCO tutorial).

Why

The notebook hard-coded three local paths that only exist on the original authors' machines:

• local/data/datasets/TACObboxSkewedResized.yaml
• local/data/datasets/TACOskewedBBoxResized
• local/data/models/best.torchscript

with no download links or preparation steps, so new contributors could not run the tutorial at
all (reported in #634).

Changes

• Automated asset download (new cells): a cross-platform, stdlib-only cell downloads and
  extracts two small public GPL-3.0 assets into local/data/ on first run, skipping when already
  present:
  • COCO128 dataset (128 images, 80 classes, YOLO format with coco128.yaml) — ~7 MB,
    from ultralytics/assets v0.0.0
  • YOLOv8n pretrained weights — ~6 MB, from ultralytics/assets v8.3.0
  • Replaces the Unix-only !mkdir -p calls with os.makedirs(..., exist_ok=True) (mirrors Fix Windows-incompatible mkdir and add COCO auto-download to tutorial notebook #613)
• TorchScript export step (new cell): one-time ultralytics export of yolov8n.pt to
  yolov8n.torchscript with dynamic=True, with a clear ImportError hint to
  pip install ultralytics if the package is missing. Pre-exported YOLO-NAS checkpoints were not
  usable because they emit tuple outputs that the repo's parser does not accept.
• Model configuration: adds "resize": {"closest_divisor": 32} so images are padded to a
  multiple of the YOLO stride (the exported dynamic model rejects other sizes); keeps
  model_format: "yolo".
• Dataset/evaluation: the dataset and model cells now consume the variables set by the setup
  cells instead of re-declaring private paths; the inference and evaluation cells use the val
  split because COCO128 has no separate test split.
• Stale outputs cleared: the stored cell outputs (1500 samples / 28 classes / AUC-PR 0.942)
  belonged to the private TACO assets and would be misleading next to COCO128.

Testing

Validated the exact notebook flow with the repo's own code (YOLODataset,
TorchImageDetectionModel.predict, .eval) on the downloaded assets:

• dataset loads: 128 samples / 80 classes, split {'val': 128}
• single-image inference works, e.g. bowl 0.93, broccoli 0.76 on a 640×480 image
• full evaluation (128 images, CPU, ~30 s) returns
  {'metrics_df', 'metrics_factory'} with mAP@[0.5:0.95] ≈ 0.300 and AUC-PR ≈ 0.330
• all notebook cells pass ast.parse; no stale TACO references remain

Fixes #634