Add generation_config.json (temperature=1.0, top_p=0.95, ZAI defaults)

Restores the base model's sampling defaults so serving engines with model-based sampling defaults apply top_p=0.95 when clients omit it. Fixes intermittent thinking-mode repetition loops / empty content reported in sgl-project/sglang#29512.

generation_config.json

@@ -0,0 +1,12 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": [
+    154820,
+    154827,
+    154829
+  ],
+  "pad_token_id": 154820,
+  "temperature": 1.0,
+  "top_p": 0.95,
+  "transformers_version": "5.12.0"
+}