But this would be either over-counting if some CI scripts download the version every once a while, or under-counting if some organization put the image on their own privately maintained mirror which is quite common.
Telemetry of any type usually fails to measure precisely the thing you want, but something adjacent that correlates strongly. What you mention are clear problems with inferring usage from downloads, but if you can infer the percentage of downloads that correlates to a machine running Fedora, you don't need much more precision.