Correctly handle duplicates in unordered interval matching

[romseygeek]

Currently, unordered interval matching does not check for duplicates,
which means that a query for to be or not to be can match a document
that contains the phrase to be or not, because the second to be matches
at the same position as the first and the AND interval algorithm does not
check for overlaps. This is counter-intuitive.

This commit adds a check to the interval builder, such that if it finds duplicates
when combining sources into an unordered AND, it combines those duplicates
into an ORDERED interval first; so to be or not to be becomes
UNORDERED(ORDERED(to, to), ORDERED(be, be), or, not)

[elasticmachine]

Pinging @elastic/es-search (:Search/Search)

[jimczi]

LGTM

[romseygeek]

This really needs to be handled in lucene, as this solution doesn't correctly handle internal gaps in intervals with repeats. I've opened https://github.com/apache/lucene-solr/pull/1097/files