Add global lock for torch.onnx.export() (#4659)

ml-agents/mlagents/trainers/torch/model_serialization.py

     This implementation is thread safe.
     """
 
+    # local is_exporting flag for each thread
+    # global lock shared among all threads, to make sure only one thread is exporting at a time
+    _lock = threading.Lock()
+
+        self._lock.acquire()
+        self._lock.release()
 
     @staticmethod
     def is_exporting():