Efficient Serving of SpeechLMs with VoxServe