fix(images): vendor spacy zh model + nltk data for air-gapped GitHub builds

- runtime: install zh_core_web_sm from vendored wheel (reused from runtime-ray)
  instead of 'python -m spacy download' (GitHub 403 on server)
- runtime / runtime-ray / backend-python: replace nltk.download (GitHub +
  nltk_data.nltk.org unreachable from server) with vendored zips extracted
  to /usr/local/nltk_data/packages/{tokenizers,taggers}

scripts/images/backend-python/Dockerfile

@@ -58,8 +58,9 @@ RUN --mount=type=cache,target=$POETRY_CACHE_DIR \
 # 注意：使用 [default] 确保包含 job submission client
 RUN pip install "ray[default]>=2.9.0"
 
-# 下载 NLTK 数据
-RUN python -c "import nltk; nltk.download(['punkt_tab','averaged_perceptron_tagger_eng'], download_dir='/usr/local/nltk_data')"
+# 下载 NLTK 数据（离线：使用仓库内 vendor 的数据包，服务器无法访问 GitHub）
+COPY scripts/images/common/nltk-data/ /tmp/nltk-data/
+RUN python -c "import zipfile, os; os.makedirs('/usr/local/nltk_data/packages/tokenizers', exist_ok=True); os.makedirs('/usr/local/nltk_data/packages/taggers', exist_ok=True); zipfile.ZipFile('/tmp/nltk-data/punkt_tab.zip').extractall('/usr/local/nltk_data/packages/tokenizers'); zipfile.ZipFile('/tmp/nltk-data/averaged_perceptron_tagger_eng.zip').extractall('/usr/local/nltk_data/packages/taggers')"
 ENV NLTK_DATA=/usr/local/nltk_data
 
 # 复制应用代码