Preconditioned Data Sparsification for Big Data with Applications to PCA and K-means