Ensure nested documents have consistent version and seq_ids

[s1monw]

Today we index dummy values for seq_ids and version on nested documents.
This is on the one hand trappy since users can request these values via
inner hits and on the other hand not necessarily good for compression since
the dummy value will likely not compress well when seqIDs are lowish.

This change ensures that we share the same field values for all documents in a
nested block. This won't have any overhead, in-fact it might be more efficient since
we even reduce the work needed slightly.

[bleskes]

LGTM. We copied this pattern from the version field:

        for (int i = 1; i < context.docs().size(); i++) {
            final Document doc = context.docs().get(i);
            doc.add(new NumericDocValuesField(NAME, 1L));
        }

Since the common version value is 1, I guess this works good for compression in the _version case but doesn't work well for seq nos. I presume maintaining the semi linear increase pattern of seq# is indeed a good thing for compression but only @jpountz knows what black magic is done there.

[martijnvg]

LGTM

[jasontedor]

LGTM.

[jpountz]

I presume maintaining the semi linear increase pattern of seq# is indeed a good thing for compression but only @jpountz knows what black magic is done there.

For the record, this did not matter before Lucene 7.0 since the random-access API made it hard to compress things efficiently, so we encoded entire segments with the same number of bits per value. This is changing as of 7.0 and we now perform block-based compression if it proves to save space.