integrate EGSteal into PyGIP (PyG + TU datasets)

[gabmun19]

📋 Summary

This PR integrates the EGSteal explanation-guided surrogate (model extraction) attack into PyGIP’s BaseAttack framework for PyTorch Geometric (PyG) workflows.

In addition, it introduces a lightweight TU graph dataset wrapper for graph classification datasets (e.g., NCI109, AIDS, Mutagenicity) and improves import robustness by making DGL-dependent components optional, enabling PyG-only usage on platforms where DGL is unavailable (e.g., macOS).

---

What Changed

New Attack: EGStealAttack

• Implemented under pygip/models/attack/egsteal/
• Supports explanation-guided extraction using:
  • CAM
  • GradCAM
  • Gradient-based explanations
  • (Optional: GNNExplainer / PGExplainer when enabled)
• Pipeline:
  • Train target (victim) GNN
  • Query target for predictions and explanations
  • Train surrogate with explanation alignment
• Reports:
  • Surrogate test accuracy and AUC
  • Target–surrogate fidelity
  • Explanation rank correlation (Kendall τ)

---

New Dataset Wrapper: TUGraph (PyG-only)

• Enables TU graph classification datasets via PyG:
  • NCI109
  • AIDS
  • Mutagenicity
• Designed for api_type="pyg"
• Cleanly integrates with existing PyGIP dataset abstractions

---

Import Robustness (DGL Optional)

• DGL-dependent datasets and attacks are now imported conditionally
• Allows:
  • Importing pygip.datasets and pygip.models.attack
  • Running PyG-based attacks (including EGSteal)
• Prevents import-time failures in environments without DGL (e.g., macOS)

---

🧪 Related Issues

---

Platform Note (macOS & DGL)

During development, DGL could not be installed on macOS due to missing prebuilt wheels.
To address this without removing DGL support:

• All DGL-dependent attacks and datasets are conditionally imported
• PyG-only workflows continue to function normally
• Clear errors are raised only if api_type="dgl" is explicitly requested

---

✅ Checklist

• My code follows the project's coding style
• I have tested the changes and verified that they work
• I have added an example script demonstrating usage
• The PR is made from a feature branch, not main

---

🧠 Additional Context

Example command to reproduce locally:

python examples/egsteal_attack.py --dataset NCI109 --epochs 5 --mode CAM