Implementation of GNN_Fingers

[SagarSharma250]

📋 Summary
This PR introduces
GNNFingers defense framework, implementing discrete adjacency matrix optimization and clipped node feature integration while establishing a foundation for multi-graph type support. Key improvements include:

Discrete Adjacency Optimization: Top-k gradient-based edge selection with straight-through estimator

Clipped Node Feature Integration: Feature value constraints based on original data distribution

Alternating Joint Learning: Improved optimization strategy for fingerprint and univerifier co-training

Architecture Foundation: Flexible structure supporting future extension to multiple graph types

🧪 Related Issues
Implements discrete optimization strategies from GNNFingers paper

Addresses feature value stability through clipping mechanisms

Enhances training stability with alternating optimization
✅ Checklist
My code follows the project's coding style
The PR is made from a feature branch, not main

🧠 Additional Context (Optional)
Discrete Adjacency Updates (_update_adjacency_discrete):

• Top-k gradient-based edge selection (configurable via top_k_ratio)
• Straight-through estimator for differentiable discrete operations
• Bidirectional edge flipping based on gradient signs

Clipped Feature Integration:
-Dynamic value clamping based on original feature distribution
-Configurable bounds for synthetic feature generation
-Gradient-aware feature updates

Alternating Optimization (_joint_learning_alternating):
-Separate update cycles for fingerprints and univerifier
-Configurable epoch ratios for each component
-Loss-based convergence monitoring