changefeedccl: Handle node unavailable error.

Treat NodeUnavailable error as a retryable changefeed error.

NodeUnavaiable error may be returned by changefeed processors if the
node is being shutdown/drained.  However, this error return is racy.
Sometimes, the coordinator would see "rpc context cancellation" error
instead -- in those cases it would treat the error as retryable.
However, sometimes it is possible to get this error propagated
(for example: when server shutdown races with starting up kv feed,
which uses Stopper, whcih may return NodeUnavailable error if it's
being shutdown).

Release Notes (bug fix): Treat node unavailable error as a retryable
changefeed error.

pkg/ccl/changefeedccl/changefeedbase/BUILD.bazel

@@ -15,6 +15,7 @@ go_library(
         "//pkg/clusterversion",
         "//pkg/jobs/joberror",
         "//pkg/jobs/jobspb",
+        "//pkg/roachpb",
         "//pkg/settings",
         "//pkg/sql",
         "//pkg/sql/catalog",

pkg/ccl/changefeedccl/changefeedbase/errors.go

@@ -14,6 +14,7 @@ import (
 	"strings"
 
 	"github.com/cockroachdb/cockroach/pkg/jobs/joberror"
+	"github.com/cockroachdb/cockroach/pkg/roachpb"
 	"github.com/cockroachdb/cockroach/pkg/sql/flowinfra"
 	"github.com/cockroachdb/errors"
 )
@@ -117,7 +118,9 @@ func IsRetryableError(err error) bool {
 		return true
 	}
 
-	return (joberror.IsDistSQLRetryableError(err) || flowinfra.IsNoInboundStreamConnectionError(err))
+	return (joberror.IsDistSQLRetryableError(err) ||
+		flowinfra.IsNoInboundStreamConnectionError(err) ||
+		errors.HasType(err, (*roachpb.NodeUnavailableError)(nil)))
 }
 
 // MaybeStripRetryableErrorMarker performs some minimal attempt to clean the