Turning Big data into tiny data: Constant-size coresets for k-means, PCA and projective clustering