Implement stage-aware real-run training pipeline

Update model configuration and training scripts with new vision backbone support and dependencies

Upload vil_dlm_model.py