Beyond Euclidean: Dual-Space Representation Learning for Weakly Supervised Video Violence Detection