sync: reland scheduler ownership refactoring, now with less crash

sync/engine/sync_scheduler_impl.cc

Index: sync/engine/sync_scheduler_impl.cc
diff --git a/sync/engine/sync_scheduler_impl.cc b/sync/engine/sync_scheduler_impl.cc
index 4d1f9a8fe99c0fb1f1f07e51ce2c897a187bccf5..c270fc525f31819f193813d70cb87614756a3345 100644
--- a/sync/engine/sync_scheduler_impl.cc
+++ b/sync/engine/sync_scheduler_impl.cc
@@ -1021,6 +1021,7 @@ void SyncSchedulerImpl::DoCanaryJob(scoped_ptr<SyncSessionJob> to_be_canary) {
   to_be_canary->GrantCanaryPrivilege();
 
   if (to_be_canary->purpose() == SyncSessionJob::NUDGE) {
+    DCHECK_EQ(pending_nudge_->session(), to_be_canary->session());
     // TODO(tim): We should be able to remove this...
     scoped_ptr<SyncSession> temp = CreateSyncSession(
         to_be_canary->session()->source()).Pass();
@@ -1048,6 +1049,9 @@ scoped_ptr<SyncSessionJob> SyncSchedulerImpl::TakePendingJobForCurrentMode() {
     pending_nudge_ = candidate.get();
     unscheduled_nudge_storage_.reset();
   }
+  // If we took a job and there's a wait interval, we took the pending canary.
+  if (candidate && wait_interval_)
+    wait_interval_->timer.Stop();

[rlarocque, 2012/10/30 17:50:20]

What if the timer fires somewhere between this line and the previous one? 
Should you use StopAndAbandon() here, or is that handled in some other way?

[tim (not reviewing), 2012/10/30 17:58:29]

Everything happens on the same message loop, so if the time logically expires
between the calls the actual callback won't fire until the stack is unwound.  At
that point, the timer won't be running, so it will not Run() any user tasks.
Because we use a OneShotTimer, Stop() also synchronously deletes the user task
(our SyncSessionJob), which means, for example, we don't have to worry about
someone resetting the timer. 

   return candidate.Pass();
 }