[CVPR 2024] NARUTO: Neural Active Reconstruction from Uncertain Target Observations