[server] only discard stale contexts that no live context depend on

[kLabz]

Since #12807, when a context is not used for 10+ minutes, it is discarded.

This led to some Could not find dependency X of Y in the cache compiler failures: in not so rare cases, the macro context is not accessed for 10+ minutes while the main context is. When this happens, next full request will try to pull some dependencies from macro context, that don't exist any more.

This PR:

• adds a way to configure stale_context_max_age_seconds (mostly for tests)
• keeps track of dependencies between contexts
  • Note: change add_context_dep to () in order to make the test fail with current development behavior
  • Also note that these dependencies are not cleared; which I think is fine, given that such dependencies are usually very stable (main <-> macro context), and avoid mistakes/overhead when trying to clear it
• only drops a stale context if no live context depends on it; meaning a stale macro context will only be dropped if its "sibling" context is also stale

The actual way to track those dependencies and read them in remove_stale_contexts is open to improvements; I asked Claude to come up with this part

Closes #12874

[Simn]

This makes sense to be conceptually. The term "sibling" alerts me a little because the macro context is closer to a child than a sibling. I think the "main" context (we still don't have a good word for that) should be considered the root, and then the macro and core contexts are its children. I'm not sure how exactly to represent that though.

[kLabz]

Yeah, I've been thinking about that too; maybe we could add them to the "main" context (as some new concept of child contexts?) upon creation directly?

[kLabz]

I switched to a parent/child relationship between contexts, which seem much cleaner.

It's not included here, but I also ran a modified version with an assert in cache_module that made sure any module dependency to another context was to a child context (had to register macro context as child before checking the modules). All server tests passed without triggering that, so it seems this is not missing some edge case.

[RblSb]

I think it doesn't crash for me too at #12846 (comment)
But class/field completion is much slower in this branch than in development for me.
Regression is not in first commit, but in second one: 76b59db
This is because staleContextMaxAge is 0 by default.

[kLabz]

Right.. fixed that. Thanks!