Identifying and typifying demographic unfairness in phoneme-level embeddings of self-supervised speech recognition models