feat: refactor dual-track architecture (Phase 1-5)

## Backend Changes
- **Service Layer Refactoring**:
  - Add ProcessingOrchestrator for unified document processing
  - Add PDFTableRenderer for table rendering extraction
  - Add PDFFontManager for font management with CJK support
  - Add MemoryPolicyEngine (73% code reduction from MemoryGuard)

- **Bug Fixes**:
  - Fix Direct Track table row span calculation
  - Fix OCR Track image path handling
  - Add cell_boxes coordinate validation
  - Filter out small decorative images
  - Add covering image detection

## Frontend Changes
- **State Management**:
  - Add TaskStore for centralized task state management
  - Add localStorage persistence for recent tasks
  - Add processing state tracking

- **Type Consolidation**:
  - Merge shared types from api.ts to apiV2.ts
  - Update imports in authStore, uploadStore, ResultsTable, SettingsPage

- **Page Integration**:
  - Integrate TaskStore in ProcessingPage and TaskDetailPage
  - Update useTaskValidation hook with cache sync

## Testing
- Direct Track: edit.pdf (3 pages, 1.281s), edit3.pdf (2 pages, 0.203s)
- Cell boxes validation: 43 valid, 0 invalid
- Table merging: 12 merged cells verified

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

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

backend/app/services/service_pool.py

@@ -14,6 +14,7 @@ from enum import Enum
 from typing import Any, Dict, List, Optional, TYPE_CHECKING
 
 from app.services.memory_manager import get_model_manager, MemoryConfig
+from app.services.memory_policy_engine import get_memory_policy_engine
 
 if TYPE_CHECKING:
     from app.services.ocr_service import OCRService
@@ -263,10 +264,16 @@ class OCRServicePool:
 
             # Clean up GPU memory after release
             try:
-                model_manager = get_model_manager()
-                model_manager.memory_guard.clear_gpu_cache()
-            except Exception as e:
-                logger.debug(f"Cache clear after release failed: {e}")
+                # Prefer new MemoryPolicyEngine
+                engine = get_memory_policy_engine()
+                engine.clear_cache()
+            except Exception:
+                # Fallback to legacy model_manager
+                try:
+                    model_manager = get_model_manager()
+                    model_manager.memory_guard.clear_gpu_cache()
+                except Exception as e:
+                    logger.debug(f"Cache clear after release failed: {e}")
 
             # Notify waiting threads
             self._condition.notify_all()