Debiasing Recommendation by Learning Identifiable Latent Confounders