Add clip to export and make optional in policy

ml-agents/mlagents/trainers/policy/torch_policy.py

             conditional_sigma=self.condition_sigma_on_obs,
             tanh_squash=tanh_squash,
         )
+        self._clip_action = not tanh_squash
         # Save the m_size needed for export
         self._export_m_size = self.m_size
         # m_size needed for training is determined by network, not trainer settings
                 vec_obs, vis_obs, masks=masks, memories=memories
             )
 
-        clipped_action = torch.clamp(action, -3, 3) / 3
+        if self._clip_action:
+            clipped_action = torch.clamp(action, -3, 3) / 3
+        else:
+            clipped_action = action
         run_out["pre_action"] = ModelUtils.to_numpy(action)
         run_out["action"] = ModelUtils.to_numpy(clipped_action)
         # Todo - make pre_action difference

ml-agents/mlagents/trainers/torch/networks.py

             self.distribution = MultiCategoricalDistribution(
                 self.encoding_size, self.action_spec.discrete_branches
             )
+        # During training, clipping is done in TorchPolicy, but we need to clip before ONNX
+        # export as well.
+        self._clip_action_on_export = not tanh_squash
 
     @property
     def memory_size(self) -> int:
             action_out = torch.stack(action_list, dim=-1)
         else:
             action_out = torch.cat([dist.all_log_prob() for dist in dists], dim=1)
+        if self._clip_action_on_export:
+            action_out = torch.clamp(action_out, -3, 3) / 3
         return (
             action_out,
             self.version_number,