Because a single abi such as x86_64 can have multiple image variants which are not distinguished by checking only the
first tag. Android API 36x were particularly affected as they have both 4kb and 16kb memory page images.
Remove the inconsistent use of `tag` to determine the image path and ID. Instead the path is uniquely determined from
the source XML metadata's location while the `id` uses part of Google's `path` ID.
The `repo.json` is updated incrementally and the last version which was not affected by these issues was `a137f263af92`
so it has been regenerated based on that version.
This regression happened during refactoring of mkrepo.rb. Do another
round of it with a focus on fixing the autoupdate and use of custom
repository XMLs.
Add an example to check this behavior (but don't run it in the test
suite, since it requires IFD).
