Post-Training Embedding Alignment for Decoupling Enrollment and Runtime Speaker Recognition Models