Sublinearly Structured Deep Neural Networks Achieve Feature Learning Consistency for Compositional Functions