Consider sharing Semantics caches between requests

Lots of file wide IDE features will re-populate the Semantics caches all the same, doing a bunch of repeated work due to us creating a new Semantics instance per request. That is kind of wasted as all the data will be the same!

So ideally we would stash away a Arc<Semantics> in GlobalState (or probably Analysis) somewhere, replace the RefCells within with RwLocks and clone the arc into requests on demand. And whenever a db invalidation comes in, we just replace the stored Arc with an Arc::default again to refresh it. This should generally speed up a couple of things I believe. That also means we should be able to cache some more stuff in Semantics that may be re-used between requests that is prone to invalidations.

Now, we can't really stash away Semantics itself as it won't be send due to the database handle within, but we can do the same with its caches directly, that is store Arc<RwLock<SourceToDefCache>> etc...

Start by locating Semantics, GlobalState, and Analysis, then trace how requests create Semantics and how database invalidations are handled. Review the existing RefCell-based caches and determine the affected cache types; done means reusable cache state is shared across requests and refreshed when invalidation occurs.