parallelize actor critic properly

ml-agents/mlagents/trainers/models_torch.py

         fut = torch.jit._fork(
             self.network_body, vec_inputs, vis_inputs, memories, sequence_length
         )
-        embedding, memories = torch.jit._wait(fut)
+        embedding, memories = torch.jit._wait(fut)
         return embedding, value_outputs, memories
 
     @torch.jit.ignore