Efficient cross-validation for kernelized least-squares regression with sparse basis expansions