Coresets for clustering in Euclidean spaces: importance sampling is nearly optimal