k8s_tagger: Service metadata is not removed from cache after Pods are deleted

[swiatekm]

We use Endpoints resources to track Service to Pod relationships and add service metadata to resource attributes. This metadata is only removed from cache when the respective Endpoints resource is deleted. However, this resource has a lifetime equal to the lifetime of the Service resource itself, and removing Pods from the Service is accomplished by updating the Endpoints resource. We don't delete any data from our cache on Endpoints updates, and so the cache entries for deleted Pods stay in the cache potentially indefinitely.

As a result, we have a slow memory leak that is only really apparent in large clusters over long time periods.