Estimating divergence functionals and the likelihood ratio by penalized convex risk minimization