feat: add RetinaNet support

[GeorgePearse]

Summary

• Implemented SingleStageDetector and RetinaNet detector.
• Implemented RetinaHead with multi-level anchor-based dense prediction.
• Ported FocalLoss from MMDetection (PyTorch version).
• Fixed BaseDenseHead to support dictionary-based test_cfg for better compatibility.
• Added forward and backward pass tests for RetinaNet.

[github-actions]

🤖 Multi-Model Consensus Review

Model	Status
Claude Sonnet 4	✅
GPT-4o	✅
Gemini 2.0 Flash	✅

✅ tests/test_models/test_forward.py

Pass Rate: 3/3 models

✅ tests/test_models/test_retinanet.py

Pass Rate: 3/3 models

✅ visdet/models/dense_heads/__init__.py

Pass Rate: 3/3 models

✅ visdet/models/dense_heads/base_dense_head.py

Pass Rate: 3/3 models

✅ visdet/models/dense_heads/retina_head.py

Pass Rate: 3/3 models

✅ visdet/models/detectors/__init__.py

Pass Rate: 3/3 models

✅ visdet/models/detectors/retinanet.py

Pass Rate: 3/3 models

✅ visdet/models/detectors/single_stage.py

Pass Rate: 3/3 models

✅ visdet/models/losses/__init__.py

Pass Rate: 3/3 models

✅ visdet/models/losses/focal_loss.py

Pass Rate: 3/3 models

[github-actions]

✅ Skylos Scan: No dead code or security issues detected.