Add topic names for coherence scores (fix #1016)

[JeanPaulShapo]

Currently there is inappropriate use of topic_name field for labeling coherence scores.
While coherence field of TopTokensScore has one number per requested topics, topic_name field has at most num_tokens numbers per requested topics.

We may true to use some parts of topic_name filed, e.g. we can try to get each num_tokens-th element, but there are at least two problems:

• It complicates code
• In fact it's wrong: actual number of gathered top tokens per topic can be less than num_tokens, because tokens with small probability cannot be gathered.

Instead of this approach, let's simply add another field into protobuf message TopTokensScore which contains topic_names and is aligned with coherence values.

[ofrei]

@JeanPaulShapo I think that's OK in this particular case - can't think of cases where serializing TopTokensScore is critical. I though it's good to be careful about protobuf serialization (particularly in messages related to Batches, Dictionaries and MasterModelConfig - TopTokensScore is none of those).