Add PyTorch CNN classifier and training pipeline

[coqylight]

Summary

• add a PyTorch CnnClassifier that mirrors the TensorFlow inference stack and exposes probability and training helpers
• load TFRecord datasets into tensor dataloaders and add a metric-aware PyTorch training loop for the CNN path
• update the training entrypoint to drive the new PyTorch implementation while keeping the existing TF-based flows for other models

Testing

• python -m compileall aav/model_training/cnn_torch.py aav/model_training/train.py

---

Codex Task