Normalize reward-to-go in C++ actor-critic (pytorch#33550)

Summary:
Comparing to the [Python implementation](https://github.com/pytorch/examples/blob/master/reinforcement_learning/actor_critic.py), it seems like the tensor of normalized reward-to-go is computed but never used. Even if it's just an integration test, this PR switches to the normalized version for better convergence.
Pull Request resolved: pytorch#33550

Differential Revision: D20024393

Pulled By: yf225

fbshipit-source-id: ebcf0fee14ff39f65f6744278fb0cbf1fc92b919

Author: nicolov

test/cpp/api/integration.cpp

@@ -193,10 +193,10 @@ TEST_F(IntegrationTest, CartPole) {
     std::vector<torch::Tensor> policy_loss;
     std::vector<torch::Tensor> value_loss;
     for (auto i = 0U; i < saved_log_probs.size(); i++) {
-      auto r = rewards[i] - saved_values[i].item<float>();
-      policy_loss.push_back(-r * saved_log_probs[i]);
+      auto advantage = r_t[i] - saved_values[i].item<float>();
+      policy_loss.push_back(-advantage * saved_log_probs[i]);
       value_loss.push_back(
-          torch::smooth_l1_loss(saved_values[i], torch::ones(1) * rewards[i]));
+          torch::smooth_l1_loss(saved_values[i], torch::ones(1) * r_t[i]));
     }
 
     auto loss =