Fix concurrent access to work_done flag.

This should have been fine without the atomic property (as it is only a single writer and multiple readers to a single bit). Nevertheless we fixed the 'race' to remove a warning in the thread sanitizer.

lib/pls/include/pls/internal/scheduling/scheduler.h

@@ -37,7 +37,7 @@ class scheduler {
   base::barrier sync_barrier_;
 
   task *main_thread_root_task_;
-  bool work_section_done_;
+  std::atomic<bool> work_section_done_;
 
   bool terminated_;
  public: