Learning Decomposable and Debiased Representations via Attribute-Centric Information Bottlenecks