kvserver: return NodeUnavailable in deferred intent resolution

When batched intent resolution races with a node shutdown, there could
be two different outcomes based on the ordering of events: (1) if the
node starts quiescing before the start of the intent resolution task,
the stopper itself returns `NodeUnavailable`; however, (2) if the
intent resolution batch is sent before the quiescing starts, the intent
resolver returns a different error (`stopping`). The `NodeUnavailable`
error is handled by the `DistSender` and retried at another node, while
the custom `stopping` error is returned to the client.

When intent resolution is async (e.g. as part of cleaning up after a
finalized transaction), the `stopping` error doesn't reach the client.
When intent resolution syncrochous (e.g. part of a transaction push),
this error can surface to the client, but only in the case of multiple
deferred intents; point-key intent resolution takes a shortcut and
doesn't use the batcher.

This commit brings parity to the error handling for both sides of the
race: they both return `NodeUnavailable`, and let the `DistSender`
retry.

Informs: #64828

Release note: None

Author: miraradeva

pkg/kv/kvserver/intentresolver/intent_resolver.go

@@ -1118,7 +1118,7 @@ func (ir *IntentResolver) resolveIntents(
 		case <-ctx.Done():
 			return kvpb.NewError(ctx.Err())
 		case <-ir.stopper.ShouldQuiesce():
-			return kvpb.NewErrorf("stopping")
+			return kvpb.NewError(&kvpb.NodeUnavailableError{})
 		}
 	}
 	return nil