Add a response_template for parsing assistant responses

Adds a `response_template` to `tokenizer_config.json` so `tokenizer.parse_response(...)` turns raw generated text back into a message dict (transformers >= 5.13, https://github.com/huggingface/transformers/pull/45847):

```python
>>> tokenizer.parse_response("A cat sitting on a mat.<|END_OF_TURN_TOKEN|>", prefix=prompt)
{'role': 'assistant', 'content': 'A cat sitting on a mat.'}
```

The model emits plain content between `<|CHATBOT_TOKEN|>` and `<|END_OF_TURN_TOKEN|>` — no thinking or tool-call fields — so the template is a single `content` field.

Verified: round-trips `apply_chat_template` -> `parse_response` at the token level, and every truncated prefix of a response parses without error (streaming-safe). No other behavior changes.

tokenizer_config.json

@@ -325,6 +325,18 @@
   "pad_token": "<PAD>",
   "padding_side": "right",
   "processor_class": "CohereCompassProcessor",
+  "response_template": {
+    "defaults": {
+      "role": "assistant"
+    },
+    "start_anchor": "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>",
+    "fields": {
+      "content": {
+        "close_pattern": "<\\|END_OF_TURN_TOKEN\\|>\\s*",
+        "content": "text"
+      }
+    }
+  },
   "sp_model_kwargs": {},
   "spaces_between_special_tokens": false,
   "tokenizer_class": "CohereTokenizer",