Scalable Importance Sampling in High Dimensions with Low-Rank Mixture Proposals