Bridging Disentanglement with Independence and Conditional Independence via Mutual Information for Representation Learning