fix(zrpc): aggregate kube resolver addresses by EndpointSlice

[artisth]

What is changed

This PR changes the zRPC Kubernetes resolver to keep endpoint addresses per EndpointSlice and aggregate all slices before updating gRPC resolver state.

Previously, every EndpointSlice update replaced the global endpoint address set with only the addresses from the updated slice. If one slice was updated to an empty endpoint list while other slices still had ready endpoints, the resolver could publish an empty address list and gRPC would report produced zero addresses.

The new implementation:

• tracks addresses by EndpointSlice identity
• aggregates addresses across all tracked slices on update
• removes only the deleted slice on delete events
• supports delete tombstones from the informer cache

Why

Kubernetes services can have multiple EndpointSlice objects. A single EndpointSlice update should not discard addresses from other slices.

Tests

• go test ./zrpc/resolver/internal/kube ./zrpc/resolver/internal
• go test ./zrpc/...