SOLR-18418: Fix complement()/intersect() with asymmetric on=

[dsmiley]

complement() and intersect() streaming expressions silently returned wrong results whenever their on= clause mapped two differently-named fields (e.g. on="_parent_document_id=document_id"), with no exception raised.

Root cause: both decorators compared tuples across streamA/streamB using streamA.getStreamSort() - a FieldComparator whose left and right field names are both streamA's own field. Applied to a streamB tuple, this always read a missing field as null, and FieldComparator's null-handling branch returns a constant, non-negative result. The on= mapping was honored by eq.test() but never by the comparison that drives the merge, so a non-matching pair was never recognized as "streamA's value is less", and instead of advancing streamB it discarded it, one tuple at a time. The first streamA value absent from streamB therefore drained streamB to EOF, after which every remaining streamA tuple hit the "streamB is EOF" branch: complement() emitted its entire input and intersect() emitted nothing.

Fix: reuse the equalitor-derived comparator that innerJoin/leftOuterJoin/ fullOuterJoin already build correctly (BiJoinStream.createIterationComparator, now promoted to StreamEqualitor.deriveComparator so BiJoinStream, Complement- Stream, and IntersectStream share one implementation). Also:

• Added StreamEqualitor.isDerivedFromLeft/isDerivedFromRight so Complement/ IntersectStream's precondition check validates each stream against the correct side of an asymmetric on=, instead of the old isDerivedFrom(), which OR's the two field checks together and can't detect an asymmetric mismatch.
• ComplementStream/IntersectStream deduped streamB with the full (asymmetric) equalitor, which compared streamB tuples using streamA's field name and so never matched; fixed via StreamEqualitor.deriveRightEqualitor(eq).
• Added StreamEqualitor.assertFieldsPresent(), called at the cross-stream comparison site, to fail loudly if an on= field is entirely absent from a tuple (a wiring bug) rather than silently treating it as null.
• Added regression tests using an asymmetric on= where streamA's first value is absent from streamB (the condition that drains streamB), asserting exact membership plus the |complement| + |intersect| == |streamA| invariant, and a focused test for the streamB dedup fix.
• Updated the complement/intersect ref guide sections with the sort/on= precondition and the sanity-check invariant.

https://issues.apache.org/jira/browse/SOLR-18418

diagnosed and solution/fix written entirely with AI

[dsmiley]

This check is going to be on every tuple reaching this point (and same for other call-sites in this PR). It has some overhead (hashmap lookups) but it saves frustration / trust issues. I think it's the right trade-off.