fix the runtimeError of in-place operation when use transformers for training

align the `input_embeds` with internvl2 original code [Line 167](https://github.com/OpenGVLab/InternVL/blob/74126371d0190a316109ddb22e4f2ca4405a7991/internvl_chat/internvl/model/internvl_chat/modeling_internvl_chat.py#L167), which may lead in a runtime error when using hf version code for training.

modeling_internvl_chat.py

@@ -101,7 +101,7 @@ class InternVLChatModel(PreTrainedModel):
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
         image_flags = image_flags.squeeze(-1)
-        input_embeds = self.language_model.get_input_embeddings()(input_ids)
+        input_embeds = self.language_model.get_input_embeddings()(input_ids).clone()
 
         vit_embeds = self.extract_feature(pixel_values)
         vit_embeds = vit_embeds[image_flags == 1]