Disentangled Latent Representation Learning for Tackling the Confounding M-Bias Problem in Causal Inference