Add metric for the number of merge buffers per query

### Description

Based on the discussion [here](https://github.com/apache/druid/pull/17360#discussion_r2054379988), we want to have a view on the count of merge buffers used by different queries, and not a point-in-time metric for that. That would help us to estimate how many merge buffers we need in total to optimize the cluster size.

### Motivation

- Mentioned above

Start with the discussion in PR #17360, especially the referenced comment, then trace where query merge-buffer usage is tracked in Apache Druid. Clarify the desired per-query metric semantics and identify relevant tests or metric documentation. Done means the metric provides a usable count of merge buffers used by queries for cluster-sizing estimates.

Written by the indexing model from the issue text.