Add document oriented collector for NRTSuggester [LUCENE-6880]

Currently NRTSuggester collects completions iteratively as they are accepted by the TopNSearcher, implying that a document can be collected more than once. In case of indexing a completion with multiple context values, the completion leads to `num_context` paths in the underlying FST for the same docId and gets collected `num_context` times, when a query matches all its contexts.

Ideally, a document-oriented collector will collect top N documents instead of top N completions by handling the docId deduplication while collecting the completions. This could be used to collect n unique documents that matched a completion query.

---
Migrated from [LUCENE-6880](https://issues.apache.org/jira/browse/LUCENE-6880) by Areek Zillur (@areek), updated Jun 08 2017
Attachments: [LUCENE-6880.patch](https://apache.github.io/lucene-jira-archive/attachments/LUCENE-6880/LUCENE-6880.patch)

Start with the NRTSuggester collector behavior described in the issue and review the attached LUCENE-6880.patch. Trace how completions are collected when multiple context values map to one docId. Done means collecting top N unique documents rather than duplicate completions, including when a query matches all contexts.

Written by the indexing model from the issue text.