The underlying storage has always returned the old object on watch
delete events when filtering. The cache watcher does not, which means a
downsteam caller gets different behavior.
This fixes the cache watcher to be consistent with our long term
behavior for watch. It may result in a behavior change (the filter
becomes more precise) but this was a regression in behavior.
Kubernetes-commit: e9e69356e4907fa4d0f45ea7e7768357ba71aba9
Clayton Coleman