feat: implement layout preprocessing backend

Backend implementation for add-layout-preprocessing proposal:
- Add LayoutPreprocessingService with CLAHE, sharpen, binarize
- Add auto-detection: analyze_image_quality() for contrast/edge metrics
- Integrate preprocessing into OCR pipeline (analyze_layout)
- Add Preview API: POST /api/v2/tasks/{id}/preview/preprocessing
- Add config options: layout_preprocessing_mode, thresholds
- Add schemas: PreprocessingConfig, PreprocessingPreviewResponse

Preprocessing only affects layout detection input.
Original images preserved for element extraction.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

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

backend/app/services/pp_structure_enhanced.py

@@ -20,6 +20,8 @@ except ImportError:
 
 import paddle
 from paddleocr import PPStructureV3
+from PIL import Image
+import numpy as np
 from app.models.unified_document import ElementType
 from app.core.config import settings
 from app.services.memory_manager import prediction_context
@@ -78,15 +80,19 @@ class PPStructureEnhanced:
         self,
         image_path: Path,
         output_dir: Optional[Path] = None,
-        current_page: int = 0
+        current_page: int = 0,
+        preprocessed_image: Optional[Image.Image] = None
     ) -> Dict[str, Any]:
         """
         Analyze document with full PP-StructureV3 capabilities.
 
         Args:
-            image_path: Path to image file
+            image_path: Path to original image file (used for cropping extracted images)
             output_dir: Optional output directory for saving extracted content
             current_page: Current page number (0-based)
+            preprocessed_image: Optional preprocessed PIL Image for layout detection.
+                               If provided, this is used for PP-Structure prediction,
+                               but original image_path is still used for cropping images.
 
         Returns:
             Dictionary with complete structure information including:
@@ -97,6 +103,8 @@ class PPStructureEnhanced:
         """
         try:
             logger.info(f"Enhanced PP-StructureV3 analysis on {image_path.name}")
+            if preprocessed_image:
+                logger.info("Using preprocessed image for layout detection")
 
             # Perform structure analysis with semaphore control
             # This prevents OOM errors from multiple simultaneous predictions
@@ -113,7 +121,16 @@ class PPStructureEnhanced:
                         'error': 'Prediction slot timeout'
                     }
 
-                results = self.structure_engine.predict(str(image_path))
+                # Use preprocessed image if provided, otherwise use original path
+                if preprocessed_image is not None:
+                    # Convert PIL to numpy array (BGR format for PP-Structure)
+                    predict_input = np.array(preprocessed_image)
+                    if len(predict_input.shape) == 3 and predict_input.shape[2] == 3:
+                        # Convert RGB to BGR
+                        predict_input = predict_input[:, :, ::-1]
+                    results = self.structure_engine.predict(predict_input)
+                else:
+                    results = self.structure_engine.predict(str(image_path))
 
             all_elements = []
             all_images = []