per-block cpu and db usage metrics are no longer reported accurately

[richvdh]

this is a regression introduced by #7120.

[richvdh]

The symptoms of this bug are basically this:

(no, we didn't magically make those processes 3 times as efficient. But we did deploy develop on 1st May.)

Those graphs chart the CPU usage used by with Measure() blocks, which are implemented via LoggingContexts. I think the basic problem is that, as of this change, the CPU usage stats are moved to the parent LoggingContext each time we switch away from the LoggingContext for the Measure block (ie, when it starts a blocking operation such as a database transaction); previously this happened only when the Measure block was all done and dusted here - ie, after Measure had added the CPU usage to its prometheus stats.

The problem is that I made that change for quite a good reason - which was that I was annoyed that LoggingContexts which did most of their work inside a child LoggingContext didn't report any CPU usage until the child finished (when it would suddenly report 5000% CPU usage).

[richvdh]

Basically, I think we need to get rid of that self._resource_usage.reset(), and instead just add the delta to the parent context. But if we do that, we need to make sure that add_database_transaction and co are propagated correctly.

[clokep]

I'm going to see if I can figure this out.

[clokep]

Done in #7426