SAM3 Video Predictor is not efficiently handling loading and memory management of frames. Should only - Load frames as tensors as needed - Discard frames and tensors when done, but only after given number of frames used by VideoPredictor to predict.