Similar to how we atomically copy log files, and other "regular" files on open we should atomically create destination data files and truncate them to their intended size. We treat the file size on disk as the canonical size but that is currently problematic if we were to crash between creating the file and truncating it to have the correct size.