Angular-margin softmax loss

Normalizes classifier weight vectors and removes the bias so softmax logits reduce to a pure angle-based score, then multiplicatively compresses each class's angular region by an integer margin m, forcing embeddings onto a hypersphere with a quantified, controllable angular margin between classes.