Disentangled Feature Learning for Real-Time Neural Speech Coding