Preserving Task-Relevant Information Under Linear Concept Removal