kv: (re-)introduce a stopgap for lack of ReplicaState synchronization

See cockroachdb#59194 and cockroachdb#58489 for more details.

In cockroachdb#58489 we observed a scary lack of synchronization around how we set
the ReplicaState for a given replica, and how we mark a replica as
"initialized". What this meant is that it was possible for the entry in
Store.mu.replicas to be both "initialized" and have an empty
ReplicaState. This is now more likely to bite us given the migrations
infrastructure attempts to purge outdated replicas at start up time
(when replicas are being initialized, and we're iterating through extan
replicas in the Store.mu.replicas map).

We believed this was addressed as part of cockroachdb#58378, but that appears not
to be the case. Lets re-introduce this stop-gap while we investigate.

Release note: None

pkg/kv/kvserver/replica.go

@@ -809,6 +809,11 @@ func (r *Replica) GetGCThreshold() hlc.Timestamp {
 
 // Version returns the replica version.
 func (r *Replica) Version() roachpb.Version {
+	if r.mu.state.Version == nil {
+		// TODO(irfansharif,tbg): This is a stop-gap for #58523.
+		return roachpb.Version{}
+	}
+
 	r.mu.RLock()
 	defer r.mu.RUnlock()
 	return *r.mu.state.Version

pkg/kv/kvserver/store.go

@@ -2719,6 +2719,10 @@ func (s *Store) PurgeOutdatedReplicas(ctx context.Context, version roachpb.Versi
 	qp := quotapool.NewIntPool("purge-outdated-replicas", 50)
 	g := ctxgroup.WithContext(ctx)
 	s.VisitReplicas(func(repl *Replica) (wantMore bool) {
+		if (repl.Version() == roachpb.Version{}) {
+			// TODO(irfansharif,tbg): This is a stop gap for #58523.
+			return true
+		}
 		if !repl.Version().Less(version) {
 			// Nothing to do here.
 			return true