1. Unless the orientation estimate that your wheel odometry provides is based on an internal imu, I would disagree with fusing the orientation over the angular velocity. Wheel odometry orientation will drift over time and throw off your yaw estimate. If your imu is calibrated properly (magnetometer) it should be able to provide you with a better orientation estimate. I'm not sure why it says that in the r_l docs, maybe I'm misinterpreting it and someone else can shed some light.
2. Regarding the high diff_drive_controller covariances: I assume the inflated covariance values correspond to Z, roll, and pitch. For a vehicle navigating in 2D, the robot would not be measuring these values. I think that in older versions of robot_localization and robot_pose_ekf, it was common practice to inflate the covariance for measurements which you did not want to include in the measurement update (high covariance means measurements will have low impact on state estimate in update). Not with robot_localization you can set in your configuration which measurements you want to fuse as outlined here.