Optimize TopGroups.merge() maxScore computation for relevance sorting

[gaobinlong]

Description

When docSort is Sort.RELEVANCE, derive each group's maxScore directly from mergedTopDocs.scoreDocs[0] instead of accumulating across shards via nonNANmax.

Also hoist docSort.equals(Sort.RELEVANCE) into a local boolean to avoid repeated equals() calls in the per-group and per-shard loops.

[benwtrent]

Do we have any benchmarks to show the improvement? It seems intuitive to me that there should be nice improvements.

[gaobinlong]

Do we have any benchmarks to show the improvement? It seems intuitive to me that there should be nice improvements.

Thanks @benwtrent , the current code in the benchmark tool luceneutil doesn't cover the TopGroups.merge() method, I cherry-picked the code in the opening PR which introduces TopGroupsCollectorManager and then did benchmark test for this optimization, here're the result:

 TaskQPS baseline      StdDevQPS my_modified_version      StdDev                Pct diff
p-value
                    TermGroup10K       43.17      (5.1%)       43.51      (7.6%)    0.8% ( -11% -   14%) 0.700
                     TermGroup1M       24.44      (6.3%)       24.79      (9.4%)    1.5% ( -13% -   18%) 0.565
                    TermGroup100      104.37      (6.9%)      108.18      (8.8%)    3.6% ( -11% -   20%) 0.145

[gaobinlong]

Hi @benwtrent, could you help to review this PR too? It's as simple as #15942.