Unfortunately, I've also found out the documentation is pretty scarce in this area - and what more - there are no high-quality open-source packages for globally referenced localization fused with local odometry. The closest you can get is https://github.com/cra-ros-pkg/robot_localization . The package allows you to fuse many sources od odometry, including absolute (GNSS). However, it has its shortcomings (more on this later). Be sure to read through its docs. There are also some very informative issues in this package.

Also, make sure to read through the other answers here: https://answers.ros.org/questions/scope:all/sort:activity-desc/tags:robot_localization,gps/page:1/ . There are some pretty good ones, although I don't have the links right now.

So what is it about the shortcomings of the robot_localization approach? Basically, it is very sensitive to having good quality GNSS data. In our case, we needed it working driving everywhere, including forests and city centers with tall buildings - both these environments can make you absolute localization jump by tens of meters (or even kilometers if you're extremely unlucky). So we're right now developing a factor-graph-based approach which would integrate the whole history (or at least its part) of measurements and find the most probable location of the robot taking this history into account.

Another thing you have to do is filter the GNSS measurements so that nonsense is not fed to your localization. It will probably be specific to your task, receiver, robot and environment. To get an idea how a complex GNSS filter could look like (from Emlid Reach M+), have a look here.