Kill exec PIDs after main container exited

Before applying this patch we killed the exec PIDs right away on
container stop which leads into the failing e2e test:

```
[sig-node] [NodeFeature:SidecarContainers] Containers Lifecycle should terminate sidecars simultaneously if prestop doesn't exit
```

This regression is now fixed by killing the exec PIDs after the main
container as well as in the same thread.

Fixes kubernetes/kubernetes#124743
Follow-up on cri-o#7937

Needs a cherry-pick since the enhancement got already backported into
supported release branches.

Signed-off-by: Sascha Grunert <sgrunert@redhat.com>

internal/oci/runtime_oci.go

@@ -849,8 +849,6 @@ func (r *runtimeOCI) StopLoopForContainer(c *Container, bm kwait.BackoffManager)
 
 	startTime := time.Now()
 
-	go c.KillExecPIDs()
-
 	// Allow for SIGINT to correctly interrupt the stop loop, especially
 	// when CRI-O is run directly in the foreground in the terminal.
 	ctx, stop := signal.NotifyContext(ctx, os.Interrupt)
@@ -935,6 +933,10 @@ func (r *runtimeOCI) StopLoopForContainer(c *Container, bm kwait.BackoffManager)
 		}
 	}, bm, true, ctx.Done())
 
+	// Kill the exec PIDs after the main container to avoid pod lifecycle regressions:
+	// Ref: https://github.com/kubernetes/kubernetes/issues/124743
+	c.KillExecPIDs()
+
 	c.state.Finished = time.Now()
 	c.opLock.Unlock()
 	c.SetAsDoneStopping()