Make the query cache a proper LRU with its own memory budget

[teh-cmc]

(Definitely not the first time this comes up, but apparently this is first issue for it.)

#7370 introduces a low-effort, low-risk safety net to prevent a degenerate situation (namely: the memory limit was exceeded even though there is nothing left to be GC'd) from getting out of hand.
I believe we need that safety net no matter what: the ability to log non-GC-able static data makes it impossible to fully prevent these kinds of degenerate cases.

Still, it is quite bad that the query cache can be hogging all the available memory budget in the first place, leading to almost all the valuable temporal user data getting GC'd in favor of old time cursors that are probably irrelevant at that point.
The query cache should have its own memory budget (probably a configurable %age of the overall memory budget), and should recycle existing cache entries when that budget is reached, like any other LRU would.

We should also investigate the possibility of not caching static data in the query-time latest-at cache: queries that return static data are much cheaper to run, and polluting the query-time cache just to point to the static tables is very wasteful.