Defending LVLMs Against Vision Attacks through Partial-Perception Supervision