Because if you want to measure the rotation, you get the cleanest readings if you can minimize the gyroscope's exposure to translational motion (mostly centripetal accelerations) due to that rotation (i.e. you want the gyro to be at the center of rotation).
A free body like an airplane, satellite, or car (maybe?) rotates about its center of gravity so you ideally would mount it there. At that location, any translational movement detected would be due to actual translation of the body (and therefore unavoidable).
But a robotic arm rotates about its joints since it is fixed at one end. So you would mount the gyro at the joint since that is what everything is rotating around. For example, if you tried to measure rotation of the elbow/forearm by mounting the gyro at the center of mass of the forearm (probably somewhere in the middle of the forearm) you would expose the gyro to translation motion due to rotation of the elbow which would make your readings less accurate. If you mounted the gyro on the elbow joint then all the translational motion detected would be due to rotation of the shoulder (unavoidable so you have to live with it).
You want to minimize the gyro's exposure to as much unnecessary translational motion as possible since those accelerations (both centripetal and linear) will distort the gyro's reading since gyros are not perfect devices.