These are purely UI/data loading considerations. It is wholly non-destructive, and deciding which tags/rating to save for the image is left to the user.

Strategy was to first extract the rating loading code from the tag loading, then simplify db vs xmp codepaths to just set values which can be compared, and finally decide what to display when we know what's up.

I feel like this code is pretty clumsy and suboptimal, but it seems to behave correctly. We may want to simplify and clean it up further, before merging to master.

Testing Note: Took me a while to realize that the XMP editor was clever enough to also update the database. This made it look like my code changes weren't working, because it still always just displayed what was set in the embedded tags. In order to force an image out of sync, I had to load up a SQLite editor and hack the appropriate rows manually.