Update checkpoint for transformers>=4.29

Following the merge of [a PR](https://github.com/huggingface/transformers/pull/24310) in `transformers` it appeared that this model was not properly converted. This PR will fix the inference and was tested using the following script:
```python
>>> from transformers import AutoTokenizer, MarianMTModel
>>> tokenizer = AutoTokenizer.from_pretrained('Helsinki-NLP/opus-mt-tc-big-en-it')
>>> inputs = tokenizer("Hey! Let's learn together", return_tensors="pt", padding=True)
>>> model = MarianMTModel.from_pretrained('Helsinki-NLP/opus-mt-tc-big-en-it')
>>> print(tokenizer.batch_decode(model.generate(**inputs)))
['<pad> Ehi, impariamo insieme</s>']
```

config.json

@@ -1,5 +1,4 @@
 {
-  "_name_or_path": "/tmp/Helsinki-NLP/opus-mt-tc-big-en-it",
   "activation_dropout": 0.0,
   "activation_function": "relu",
   "architectures": [
@@ -40,7 +39,7 @@
   "share_encoder_decoder_embeddings": true,
   "static_position_embeddings": true,
   "torch_dtype": "float16",
-  "transformers_version": "4.22.0.dev0",
+  "transformers_version": "4.34.0.dev0",
   "use_cache": true,
   "vocab_size": 54422
 }

generation_config.json

@@ -12,5 +12,5 @@
   "num_beams": 4,
   "pad_token_id": 54421,
   "renormalize_logits": true,
-  "transformers_version": "4.32.0.dev0"
+  "transformers_version": "4.34.0.dev0"
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a87b2cd9f24523d0cb95d8390bd09ad6888d1adb9fdfbe60ef13b44fc7902d96
-size 575765900
+oid sha256:5c8e2b4c7c161fb30cb609818d187101697d06e5c6265700adfa91126478b6d1
+size 464309380

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3815249ce2b14f6b2ffd3de1050b183865f70071d641011ee5de5073d3a90a64
-size 575827779
+oid sha256:d9d0637a282266f2a6c9b7a0000640a505fe85352eed703e09ff161f9273b905
+size 464366341