Unsupervised adversarial + Procrustes cross-space alignment

Learns a single orthogonal linear map between two independently-trained embedding spaces with no parallel/supervised anchor pairs, via an adversarial discriminator followed by iterative orthogonal-Procrustes refinement on the map's own most-confident pairs, then validates the map by real downstream retrieval/translation accuracy.