Choosing Public Datasets for Private Machine Learning via Gradient Subspace Distance