Share partition cache between selector instances

[MaxiBoether]

Currently, each Selector process has their own self._trigger_cache due to the introduction of gRPC using multiprocessing.Processes for true parallelism in #301.

The problem is that we cannot easily use a manager.Dict() here to share a dictionary across processes (leaving performance implications out for now). The problem is that new Selector instances get allocated at runtime, and any Selector instance would need to instantiate a new manager.Dict(). This is not possible as a manager cannot be shared with child classes.

We need to re-think how we do multiprocessing in the Selector for this, probably..

Should be addressed with #178.

#309 has the same root cause, which is requiring access to a parent manager in subclasses.