Improving the accuracy of likelihood-based inference in meta-analysis and meta-regression