Memory-Efficient Training for Deep Speaker Embedding Learning in Speaker Verification