Fix Kubernetes strategy crash on endpoint addresses without targetRef

[dajiaohuang]

Summary of changes

EndpointAddress.targetRef is optional in the Kubernetes API. Endpoints which are not backed by a pod — manually managed ones pointing at external addresses, for example — have no targetRef at all, and parsing such a response raised a FunctionClauseError in parse_response/2, which runs from init/1. The strategy then failed to start, so the node never joined the cluster.

Reproducing it needs no real cluster: the added VCR cassette returns one address with a targetRef and one without. Before the change the strategy fails to start with:

** (FunctionClauseError) no function clause matching in anonymous fn/1 in Cluster.Strategy.Kubernetes.parse_response/2
    (libcluster 3.5.0) lib/strategy/kubernetes.ex:473: ...
    (libcluster 3.5.0) lib/strategy/kubernetes.ex:250: Cluster.Strategy.Kubernetes.init/1

The fix keeps using targetRef.namespace when it is present, and otherwise falls back to the namespace the endpoints were queried from. Every address in an endpoints response belongs to that namespace, because the API request is namespace-scoped, so names built for :hostname and :dns modes stay correct.

Checklist

• New functions have typespecs, changed functions were updated (only private functions changed, so no typespec or public API is affected)
• Same for documentation, including moduledocs (no documented behaviour changed)
• Tests were added or updated to cover changes (test/kubernetes_test.exs, plus the cassette test/fixtures/vcr_cassettes/kubernetes_no_target_ref.json)
• Commits were squashed into a single coherent commit
• Notes added to CHANGELOG file which describe changes at a high-level

mix test (40 tests, 0 failures) and mix format --check-formatted both pass.