Clear cumulative_returns_since_policy_update (#2120)

Before the CSV file's mean rewards would lag much behind the rest of the code since this buffer was never cleared.

Author: Ervin T

ml-agents/mlagents/trainers/ppo/trainer.py

@@ -422,6 +422,7 @@ def update_policy(self):
             number_experiences=len(self.training_buffer.update_buffer["actions"]),
             mean_return=float(np.mean(self.cumulative_returns_since_policy_update)),
         )
+        self.cumulative_returns_since_policy_update = []
         n_sequences = max(
             int(self.trainer_parameters["batch_size"] / self.policy.sequence_length), 1
         )