Use `<|im_end|>` as EOS token in `tokenizer_config.json`

This is more consistent with `config.json` and `generation_config.json`.
Moreover, this fixes endless outputs in llama.cpp (since their
conversion script gets the EOS token from `transformers.AutoTokenizer`
which reads `tokenizer_config.json`).

tokenizer_config.json

@@ -2,7 +2,7 @@
   "add_prefix_space": false,
   "bos_token": "<|endoftext|>",
   "clean_up_tokenization_spaces": true,
-  "eos_token": "<|endoftext|>",
+  "eos_token": "<|im_end|>",
   "model_max_length": 1000000000000000019884624838656,
   "tokenizer_class": "GPTNeoXTokenizer",
   "unk_token": "<|endoftext|>"