make visual input channel first for export

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

         seq_len_dim = [1]
         dummy_vec_obs = [torch.zeros(batch_dim + [self.policy.vec_obs_size])]
         dummy_vis_obs = [
-            torch.zeros(batch_dim + list(shape))
+            torch.zeros(batch_dim + [shape[2], shape[0], shape[1]])
             for shape in self.policy.behavior_spec.observation_shapes
             if len(shape) == 3
         ]

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

 
         for idx, encoder in enumerate(self.visual_encoders):
             vis_input = vis_inputs[idx]
-            vis_input = vis_input.permute([0, 3, 1, 2])
+            if not torch.onnx.is_in_onnx_export():
+                vis_input = vis_input.permute([0, 3, 1, 2])
             hidden = encoder(vis_input)
             encodes.append(hidden)