Fine performance metrics: measure time spent entering/exiting thread pools

[crusaderky]

• Part of Fine performance metrics meta-issue #7665
• Related to Compression slows down network comms  #7655
• Related to Task should be executing upon rejoining. #5882

When investigating #7655, I got the suspicion that we may have a bottleneck caused by the fact that there's only one worker in the offload executor.
Additionally, there's a known issue (#5882) where, whenever a task calls rejoin(), you will have one less thread available then what the WorkerStateMachine believes is available, which will cause Worker.execute to get stuck for a long time on the call to run_in_executor.

In both cases, as of #7586 this time is displayed as ("execute", <prefix>, "other", "seconds").
Wrap these two calls to run_in_executor in with context_meter.meter("offload") and with context_meter.meter("executor") respectively.