kgo.Client: add explicit ShardedRequest; .Broker: add RetriableRequest

The prior Request always merged responses, which was not perfect all of
the time. If one request to a broker failed of many requests, then we
would only have a partial response, and it would not be clear that the
response was partial.

The new ShardedRequest returns _all_ issued request/response/error pairs
and the broker that the requests went to, or an unknown broker if the
request could not be issued.

This is a rather large change that has a couple of other side effects:

- the admin controller is un-cached whenever a NOT_CONTROLLER error is
returned from an admin request. Previously, we would continue using the
cached controller until a metadata reload.

- this adds sharding for DescribeConfigs, AlterConfigs, and
IncrementalAlterConfigs

- for AlterReplicaLogDirs and DescribeLogDirs, requests are issued to
all replicas, rather than just the leader

- some requests in the Request flow are now retried, rather than failing
(such as if we cannot load metadata on a certain broker when looking up
where to send the request)

Since some aspect of retrying has been abstracted into a function, this
also now easily adds retrying an individual, broker dedicated request to
the Broker type.

pkg/kgo/broker.go

@@ -63,6 +63,10 @@ type waitingResp struct {
 	err     error
 }
 
+var unknownMetadata = BrokerMetadata{
+	NodeID: -1,
+}
+
 // BrokerMetadata is metadata for a broker.
 //
 // This struct mirrors kmsg.MetadataResponseBroker.