feat: 新增外部数据源拉取与 DPO 输出格式支持

- 支持从 PostgreSQL 数据库拉取结构化数据作为训练来源
- 新增 DPO (Direct Preference Optimization) 输出类型
- 支持 chosen/rejected 字段的编辑、校验和发布
- 完善数据预处理切分逻辑和元数据管理
- 移除 OCR 扫描 PDF 功能，保持基础文本解析能力

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

backend/tests/test_data_process_store.py

@@ -1335,6 +1335,39 @@ def test_publish_rejects_invalid_reasoning_output_format() -> None:
         )
 
 
+def test_publish_dpo_writes_chosen_and_rejected_jsonl() -> None:
+    conn = _PublishConnection(
+        [
+            {
+                "id": "result-dpo",
+                "status": "valid",
+                "instruction": "如何处理扫描 PDF？",
+                "input": "",
+                "output": "仅在无文本层时执行 OCR。",
+                "chosen": "仅在无文本层时执行 OCR。",
+                "rejected": "所有 PDF 都执行 OCR。",
+                "preview_item_id": "preview-dpo",
+            }
+        ]
+    )
+
+    published = _PublishStore(conn, {"output_type": "dpo"}).publish(
+        "task-dpo",
+        {
+            "dataset_name": "偏好数据",
+            "storage_type": "local",
+            "format": "dpo",
+            "split": {"train": 100, "validation": 0, "test": 0},
+        },
+    )
+
+    record = conn.records[0]["raw"]
+    assert record["chosen"] == "仅在无文本层时执行 OCR。"
+    assert record["rejected"] == "所有 PDF 都执行 OCR。"
+    assert "output" not in record
+    assert published["datasets"][0]["metadata"]["format"] == "dpo"
+
+
 def test_source_storage_descriptor_accepts_owned_local_and_legacy_db_references() -> None:
     task_id = "dpt_task"
     source_file_id = "dpsf_source"