feat: refactor summarizer and PDF extraction pipeline

- Split summarizer into summary_generator and summary_persister modules
- Refactor pdf_image_extractor to two-phase pipeline with PicoDet layout detection
- Add layout_detector service for PicoDet-S_layout_3cls integration
- Add exceptions module with ConflictError and NotFoundError
- Improve admin dashboard with better statistics and task management
- Add design review document with system optimization suggestions
- Add new tests for crawler, pdf_downloader, pipeline, and summary_utils
- Update dependencies and configuration
- Clean up dead code and improve error handling

tests/conftest.py

@@ -161,7 +161,13 @@ def sample_summary_dict() -> dict:
         "results": {
             "main_findings": "在长文本基准 LongBench 上取得了 SOTA 结果，平均得分提升 3.2 个百分点。推理速度相比全注意力提升了 2 倍，显存占用降低 60%。在 32k 序列长度下仍保持与全注意力相当的生成质量。",
             "benchmarks": [
-                {"task": "长文本摘要", "metric": "ROUGE-L", "this_work": "42.1", "baseline": "38.9", "improvement": "+3.2"},
+                {
+                    "task": "长文本摘要",
+                    "metric": "ROUGE-L",
+                    "this_work": "42.1",
+                    "baseline": "38.9",
+                    "improvement": "+3.2",
+                },
             ],
             "limitations": "在超长文本（>100k tokens）上效果有所下降，主要原因是全局采样点数量不足以覆盖所有关键信息。此外，在小规模数据集上的优势不如大规模数据集明显。",
         },