fix deadlock when dump samples with filter

[Zenglinxiao]

Currently, when dumping samples after applying transforms in the onmt_build_vocab, we may encounter a deadlock issue if the filtertoolong exist in the transform pipe.
The multiprocessing dumping using num_threads queue, each corresponds to one shard of original corpus and they are supposed to be equal size. This equal size assumption will be broken by the filtertoolong transform as there may have a different number of examples removed from that shard which will also break the ordering when writing.
To fix it, we need to pass a placeholder even that example has been filtered. Thus guaranteed the equal size assumption.