improve default engine `copy_atomic` implementation

[zachschuermann]

for now, we do the simple+correct thing: GET then PUT the file. ideally we should try to do server side copy? but S3 can't do that..