Sibling instances on different workers would not receive invalidation
events since those are sent under the namespace of the publishing
worker, which has a different address for its mlcache instances than
others (potentially).

Instead of relying on the memory address, we now require users of
mlcache to simply name their cache. Several mlcache instances bearing
the same name are simply sharing the same data (same LRU and same
namespace in the shared dict).

In order to properly GC our instances, we must now keep track of how
many instances are using the same name and thuse, only GC an LRU
instance if no mlcache instance is using it (0 references left).