Fix crash in trainer tests (trainer_metrics)

ml-agents/mlagents/trainers/tests/test_ppo.py

     buffer.update_buffer["extrinsic_returns"] = buffer.update_buffer["rewards"]
     buffer.update_buffer["extrinsic_value_estimates"] = buffer.update_buffer["rewards"]
     trainer.training_buffer = buffer
+    # Make sure trainer_metrics doesn't crash
+    trainer.trainer_metrics.last_mean_return = 0
     trainer.update_policy()
     # Make batch length a larger multiple of sequence length
     trainer.trainer_parameters["batch_size"] = 128