Add ERNIE-4.5-21B-Thinking (Q1_0) to model registry

- Add unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF with TQ1_0 quantization
- MoE architecture: 21B total params / 3B activated per token
- 128K context window (capped at 32K for CPU performance)
- Inference settings: temp=0.7, top_p=0.8, top_k=40 (Baidu/Unsolt defaults)
- Thinking-only mode (no /think toggle needed)
- Largest model in registry (21B total parameters)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

app.py

@@ -191,6 +191,19 @@ AVAILABLE_MODELS = {
             "repeat_penalty": 1.1,
         },
     },
+    "ernie_21b_thinking_q1": {
+        "name": "ERNIE-4.5 21B Thinking (128K Context)",
+        "repo_id": "unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF",
+        "filename": "*TQ1_0.gguf",
+        "max_context": 131072,
+        "supports_toggle": False,  # Thinking-only mode
+        "inference_settings": {
+            "temperature": 0.7,
+            "top_p": 0.8,
+            "top_k": 40,
+            "repeat_penalty": 1.05,
+        },
+    },
 }
 
 DEFAULT_MODEL_KEY = "qwen3_600m_q4"