Column subset serialization uses an unnecessary -1L for large subsets

Follow up to CASSANDRA-9894. It's possible to completely remove a 9 byte overhead for large columns, as we know upfront which kind of serialization we will have done.