SOC-ICNN: From Polyhedral to Conic Geometry for Learning Convex Surrogate Functions