fix: make torch import optional and add PaddlePaddle GPU memory management

Problem:
- Backend failed to start with ModuleNotFoundError for torch module
- torch was imported as hard dependency but not in requirements.txt
- Project uses PaddlePaddle which has its own CUDA implementation

Changes:
- Make torch import optional with try/except in ocr_service.py
- Make torch import optional in pp_structure_enhanced.py
- Add cleanup_gpu_memory() method using PaddlePaddle's memory management
- Add check_gpu_memory() method to monitor available GPU memory
- Use paddle.device.cuda.empty_cache() for GPU cleanup
- Use torch.cuda only if TORCH_AVAILABLE flag is True
- Add cleanup calls after OCR processing to prevent OOM errors
- Add memory checks before GPU-intensive operations

Benefits:
- Backend can start without torch installed
- GPU memory is properly managed using PaddlePaddle
- Optional torch support provides additional memory monitoring
- Prevents GPU OOM errors during document processing

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

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

backend/app/services/pp_structure_enhanced.py

@@ -9,7 +9,16 @@ import logging
 from pathlib import Path
 from typing import Dict, List, Optional, Tuple, Any
 import json
+import gc
 
+# Optional torch import for additional GPU memory management
+try:
+    import torch
+    TORCH_AVAILABLE = True
+except ImportError:
+    TORCH_AVAILABLE = False
+
+import paddle
 from paddleocr import PPStructureV3
 from app.models.unified_document import ElementType
 
@@ -155,6 +164,18 @@ class PPStructureEnhanced:
             logger.error(f"Enhanced PP-StructureV3 analysis error: {e}")
             import traceback
             traceback.print_exc()
+
+            # Clean up GPU memory on error
+            try:
+                if TORCH_AVAILABLE and torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+                    torch.cuda.synchronize()
+                if paddle.device.is_compiled_with_cuda():
+                    paddle.device.cuda.empty_cache()
+                gc.collect()
+            except:
+                pass  # Ignore cleanup errors
+
             return {
                 'elements': [],
                 'total_elements': 0,