fix: set `clean_up_tokenization_spaces` to `false`

`clean_up_tokenization_spaces=true` causes `tokenizer.decode()` to silently strip spaces before punctuation, producing incorrect decoded text for Llama 3's BPE tokenizer. This was inherited from a HuggingFace transformers library default — Llama 2 had it set to `false`, and Llama 4 already ships with `false`.

See the full writeup with reproduction, impact analysis, and history: https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct/discussions/356

The fix is a one-line change in `tokenizer_config.json`.

tokenizer_config.json

@@ -2050,7 +2050,7 @@
     }
   },
   "bos_token": "<|begin_of_text|>",
-  "clean_up_tokenization_spaces": true,
+  "clean_up_tokenization_spaces": false,
   "eos_token": "<|end_of_text|>",
   "model_input_names": [
     "input_ids",