Optimise phrasescorer matches

[iprithv]

Description

Optimise PhraseScorer by short-circuiting non-competitive documents in TOP_SCORES mode.

In PhraseScorer.TwoPhaseIterator.matches(), the matcher.reset() call is moved after the competitive score check. This allows skipping the expensive reset() work for documents whose maximum possible score falls below minCompetitiveScore.

Why this matters

reset() is expensive because it:

• Calls PostingsEnum.freq() which decodes a full PFOR block per term (see Lucene104PostingsReader)
• For exact phrases: initialises position iteration state
• For sloppy phrases: rebuilds a priority queue via initPhrasePositions()

By checking maxFreq() first and short-circuiting when the document can't be competitive, all of this work is avoided.

Implementation

To support calling maxFreq() before reset(), both ExactPhraseMatcher and SloppyPhraseMatcher now load term frequencies eagerly inside maxFreq() and track this via a boolean freqsLoaded flag:

• When reset() sees freqsLoaded == true (maxFreq was called), it skips redundant freq() calls
• When maxFreq() was not called (non-TOP_SCORES mode, or before minCompetitiveScore is established), reset() follows the original single-loop code path with zero added overhead

The boolean flag approach avoids the cost of a virtual approximation.docID() call that would be needed with a lastDocId guard.

Testing

• All 89 phrase-related tests pass
• All org.apache.lucene.search.* tests pass
• Added TestPhraseMatcherContract: directly tests that maxFreq() can be called before reset() for both exact and sloppy matchers
• Added JMH benchmark (PhraseScorerBenchmark) with 3 forks, 5 warmup iterations, 5s measurement intervals

Benchmarking

JMH benchmark (1M docs, 0.1% phrase matches, 50% docs with terms but no phrase):

Benchmark	Baseline (main)	Candidate (this PR)	Delta
benchmarkExactTopScores	2.767 ± 0.181 ops/ms	2.907 ± 0.261 ops/ms	+5.1%
benchmarkSloppyTopScores	0.070 ± 0.002 ops/ms	0.071 ± 0.004 ops/ms	+1.4%

[github-actions]

This PR has not had activity in the past 2 weeks, labeling it as stale. If the PR is waiting for review, notify the dev@lucene.apache.org list. Thank you for your contribution!