For retained rendering, we don't want to push the offset down to each leaf layer. Otherwise, changing an offset layer on the very high level could cascade the change to too many leaves, which means that we can't retain them.

To not push the offset downwards, we simply push a TransformLayer when there's an offset. Skia has a fast path for concatenating scale/translation-only matrix so this operation should be fast (no performance regression is measured on Moto G4).

This is our first step towards https://github.com/flutter/flutter/issues/21756