feat(engine): balanced 역할 jinja thinking + checkpoint RAM 오프로드

- --jinja + --chat-template-kwargs '{"enable_thinking":true}' 추가
- -cram 8192: context checkpoint를 GPU 대신 CPU RAM에 저장
  (GPU CUDA OOM 크래시 방지 — cuMemSetAccess 실패 at device:1)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

config/engine_models.json

@@ -46,6 +46,9 @@
         "--mmproj",
         "models/mmproj-F16.gguf",
         "--no-mmproj-offload",
+        "--jinja",
+        "--chat-template-kwargs",
+        "{\"enable_thinking\":true}",
         "-ngl",
         "999",
         "-c",
@@ -63,7 +66,9 @@
         "-b",
         "512",
         "-ts",
-        "0.48,0.52"
+        "0.48,0.52",
+        "-cram",
+        "8192"
       ]
     },
     "deep-coder": {