Towards the Identifiability in Noisy Label Learning: A Multinomial Mixture Approach