Geometric Analysis of Self-Supervised Vision Representations for Semantic Image Retrieval