MatrixStatsAggregatorTests#testTwoFields fails if aggs are reduced

[polyfractal]

As part of #36020, I accidentally stumbled onto a test bug. testTwoFields() fails if it is changed to searchAndReduce() instead of search(). The variance of the test's MultiPassStats object no longer matches the reduced matrix stats response.

It's not clear if this is because the mock MultiPassStats is doing something wrong, or if there's a bug when the matrix agg's reduction is buggy. MatrixStats uses a per-shard object (RunningStats) which is eventually converted into the final response object (MatrixStatsObject), at which point the correlation is calculated. It might be part of this process that is buggy, or perhaps how the stats objects are finalized after reduction.

[elasticmachine]

Pinging @elastic/es-analytics-geo