dns-discovery-netty: fix SRV resurrection during A-record backoff

[bryce-anderson]

Motivation:

An SRV-rooted A-record resolution that fails with a transient error enters a backoff retry. That retry lived outside the publisher: the target was removed from aRecordMap for the duration of the backoff and a fresh ARecordPublisher (with empty state) was created per attempt. Discarding the per-attempt state caused two problems:

• A concurrent SRV removal during the backoff window found no publisher to cancel, so the removal was lost and the pending retry resurrected a target DNS had reported as gone.
• The last-known addresses could not be retracted when the target was removed mid-backoff, so the load balancer kept stale addresses until the whole SRV subscription was torn down.

Modifications:

• Fold the retry into AbstractDnsSubscription: on a resolution failure a backoff re-query is scheduled on the same subscription, preserving the last-known addresses across attempts. Gated by a retryResolveFailures flag, enabled only for SRV-rooted A lookups. NXDOMAIN still invalidates the addresses first; other failures keep them.
• A single ARecordPublisher now serves a target for its whole lifetime and is registered in aRecordMap up front, so a concurrent SRV removal always finds it (even before the first subscribe or mid-backoff), cancels it, and retracts its last-known addresses.
• Remove the external retryWhen wrapper, preserving its Reactive Streams Rule 2.13 wrapping of subscriber exceptions so those still revoke state.

Result:

A removed SRV target is no longer resurrected by an in-flight backoff retry, and its last-known addresses are retracted immediately on removal. A transient failure that recovers to the same addresses no longer emits a redundant AVAILABLE, matching a TTL re-query that returns an unchanged result.