Paper on masked scene modeling for pre-training in 3D accepted at CVPR 2025