Marginally Interpretable Linear Transformation Models for Clustered Observations