Physics-constrained deep learning for high-dimensional surrogate modeling and uncertainty quantification without labeled data