Ports scene extract, face ONNX preprocess, and net reference resize to Torch/CUDA with minimal CPU sync, while keeping the CPU parity path unchanged. Co-authored-by: Cursor <cursoragent@cursor.com>
