Cluster-level reduction for ES|QL

[dnhatn]

Remote clusters can be located in different regions or zones, with high latency transport requests. For instance, the latency between us-west and eu-west can exceed 100ms. When using the _search API with CCS and ccs_minimized_roundtrips=true, only one cross-cluster transport request is needed per remote cluster. However, ES|QL with CCS requires multiple cross-cluster transport requests, including field-caps requests, cluster-compute requests, and exchange requests.

I investigated an issue where a remote cluster generated 70 small pages, with a latency greater than 120ms between clusters. Given that we have three concurrent clients fetching pages, this process alone could take more than 2.8 seconds (70/3 * 120ms).

Node-level reduction should help alleviate this problem, but it is currently disabled by default and only enabled in snapshot builds. And if the remote cluster targets a large number of nodes, cluster-level reduction will be necessary to avoid latency issues. The implementation of cluster-level reduction can be similar to that of node-level reduction. However, the implementation should also consider the trade-off between reducing latency and avoiding duplicate work on both the coordinator in the querying cluster and the remote clusters.

[elasticsearchmachine]

Pinging @elastic/es-analytical-engine (Team:Analytics)

[dnhatn]

Implemented in #117731