Cherry-pick separate critic only for PPO (#4661) (#4666)

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

             behavior_spec,
             self.trainer_settings,
             condition_sigma_on_obs=False,  # Faster training for PPO
-            separate_critic=behavior_spec.action_spec.is_continuous(),
+            separate_critic=True,  # Match network architecture with TF
         )
         return policy