Slowdown when running nested statement with many partitions

(Filing based on investigation of a support case)

When a statement uses a partitioned table, there is O(#partitions_in_table) overhead even if the statement actually touches O(1) partitions.

We always had this overhead (there is MDEV-20250 to get rid of it), but in 10.6 it got larger.

In order to observe the overhead increase one needs:

• A table with a lot (thousands) or partitions.
• A statement that touches one or a few rows in one/few partitions. The statement must be sufficiently lightweight so the extra overhead is visible.
• The above statement must be invoked many times by the embedding statement. This way, the overhead adds up and can be observed.

The total slowdown observed was around 40%.

Causes

• One known reason is handler_stats_disable() and related calls which were added
  by MDEV-31558. It has unnecessary loops that go through all partitions.

Overhead that already existed in 10.5:

• Calls to extra(HA_EXTRA_.._CHILDREN). These are no-ops if one is not using MERGE engine.

• Calls to extra(HA_EXTRA_*_KEYREAD). These only make sense for the partitions that are touched by the query