Improving the Privacy and Practicality of Objective Perturbation for Differentially Private Linear Learners