Fix stats performance

[masseyke]

This fixes the N^2 performance problem described in #97222. In addition to restoring the previous partial fix (#130857), it does the following:

1. IndicesQueryCache::getStats now accepts a Supplier so that we can only call IndicesQueryCache::getSharedRamSizeForAllShards if it is absolutely needed. This fixes an N^2 performance problem that Improving statsByShard performance when the number of shards is very large #130857 introduced. If a user called TransportIndicesStatsAction but did not request query cache stats, then before Improving statsByShard performance when the number of shards is very large #130857 we did not enter the N^2 loop (it was only entered if a user did request query cache stats). But after Improving statsByShard performance when the number of shards is very large #130857, we had the N^2 performance all the time. This is a pretty big problem for clusters with large shards since this is called very frequently (including every 30 seconds by a background task).
2. It fixes the N^2 performance in TransportIndicesStatsAction by sharing state across all shardOperation calls on a single node using the new NodeContext feature from Adding NodeContext to TransportBroadcastByNodeAction #138057.

Closes #97222

[elasticsearchmachine]

Hi @masseyke, I've created a changelog YAML for you.