fix: resolve code review issues — BBOX_TENSOR type, validation params, cache safety, dead write removal
- RETURN_TYPES: "FLOAT" → "BBOX_TENSOR" for [B,6] metadata tensor (prevents scalar misinterpretation)
- Validation: remove hardcoded phantom params, pass actual scaling_method; fix field name to match Pydantic model
- YOLO cache: device-aware key ("model::device"), threading.Lock, CUDA-only torch.compile()
- Scaling validator: add "bicubic", case-insensitive normalization to match node uppercase values
- Remove dead rgba[:,:,3] write-back in GrabCutRefinement (only alpha used downstream)
- Add initial_mask.unsqueeze(0) for single-image 2D masks
- Remove unused Annotated import from validation.py
limbicnation
77b189a05a