Addressing Label Shift in Distributed Learning via Entropy Regularization