It must be awfully convenient to believe that the only possible reason for it to be on is the same reason you made up entirely based on nothing. Personally, I don't find that logic convincing.
a) Compensating for missing location metadata is a valid feature for a photo library. Peer competitor Google Photos also implements location estimation from landmarks.
b) The sizes of contemporary general vision models and the size of the vector database for matching potentially millions of landmarks suggest that this is not suitable for running on-device.
c) Apple's entire strategy is to do cloud computation without private data, so it stands to reason that they believe they're not using private data.
You asked "If not [for training data], then why the default opt-in?" and I gave three alternative reasons.