Similar Items: Exploring the Efficiency of 3D-Stacked AI Chip Architecture for LLM Inference with Voxel