The current version of NuclearJS uses a cache key consisting of store
states (monotomically incresing ID per store).  This has the
disadvantage of allowing only a single level of depth when figuring out
if a cache entry is stale.  This leads to poor performance when the
shape of a Reactor's state is more deep than wide, ie a store having
multiple responsibilities for state tracking.

The implementation is as follows:

- Consumer can set the maxCacheDepth when instantiating a reactor
- Getters are broken down into the canonical set of keypaths based on
the maxCacheDepth
- Add a keypath tracker abstraction to maintain the state value of all
tracked keypaths
- After any state change (`Reactor.__notify`) dirty keypaths are
resolved and then based on which keypaths have changed any dependent
observers are called

Instead of doing a full recursive traversal of the keypathStates tree,
keep track of all changed paths as use that as an indicator of the paths
the need to be cleaned