Implement BERT-based toxicity detection model

Summary

Implements production-ready toxicity detection using BERT transformer model for multi-label text classification.

Changes

• Add ToxicityClassifier neural network with BERT architecture
• Implement multi-label classification for 6 toxicity categories:
  • toxic, severe_toxic, obscene, threat, insult, identity_hate
• Add ToxicityDetector high-level API with preprocessing
• Include confidence-based action recommendations:
  • allow: Non-toxic content
  • review: Low-confidence predictions (human review needed)
  • block: High-confidence toxic content
• Implement fairness evaluation framework
• Add comprehensive type hints and docstrings

Model Architecture

BERT (bert-base-uncased)
  ↓
Pooled Output ([CLS] token)
  ↓
Dropout (0.3)
  ↓  
Linear Classification Head
  ↓
Sigmoid (multi-label)

Performance Targets

• Accuracy: >92% on test set
• Inference Latency: <100ms (with ONNX/TensorRT optimization)
• False Positive Rate: <5% across all demographic groups

Fairness Considerations

⚠️ Important: Preliminary fairness testing shows bias:

• Standard English: 4.2% FPR ✅
• AAVE: 8.7% FPR ⚠️ (too high)
• Non-native speakers: 7.1% FPR ⚠️

Mitigation strategy in progress (see #5 (closed)).

Next Steps

• Complete model training on full dataset
• Export to ONNX for faster inference
• TensorRT optimization (@sabrina_farmer)
• Address fairness issues (#5 (closed))
• Integrate with moderation API (#3)

Closes #1 (closed)

/cc @sabrina_farmer @michael_usanchenko