Add GNNFingers unified pipeline example (+ AUROC/ARUC, matcher options)

[Serina-2002]

Summary

• Adds a runnable example at examples/gnnfingers/gnnfingers_full_pipeline.py that:
  • trains victim models (node/graph), link-pred encoders, and an optional pairwise matcher,
  • generates positive/negative variants,
  • computes Fingerprinting AUROC and ARUC,
  • saves plots/CSVs/JSON under outputs/.
• Includes examples/gnnfingers/README.md.
• Uses PYGIP_DEVICE if set (e.g., cuda:0), otherwise falls back to CPU/GPU availability.
• Resolves --outdir relative to the repo so outputs land in PyGIP/outputs/.

Why

A single end-to-end entry point makes it easy to reproduce fingerprint-based
ownership verification experiments and compare metrics quickly, matching the
repository’s example-first structure.

How to run

# Cora node classification + fingerprinting
python examples/gnnfingers/gnnfingers_full_pipeline.py --dataset Cora --variants 200 --save-plots

# PROTEINS graph classification + optional matcher
python examples/gnnfingers/gnnfingers_full_pipeline.py --dataset PROTEINS --variants 200 \
  --train-matcher --matcher-epochs 100 --matcher-lr 5e-4 --save-plots

[Edlison]

Could you check that your code follows our implementation guideline? The new method should be under models/defense, and the examples folder should include only runnable scripts. If anything in the guideline is unclear, please let me know so we can align.