Representation Learning for Discovering Phonemic Tone Contours