feat: aggressive image optimization for PPDocLayoutV3 only

- Remove doclayout-yolo (~4.8GB, torch/torchvision/triton)
- Replace opencv-python with opencv-python-headless (~200MB)
- Strip debug symbols from .so files (~300-800MB)
- Remove paddle C++ headers (~22MB)
- Use cuda:base instead of runtime (~3GB savings)
- Simplify dependencies: remove doc-parser extras
- Clean venv aggressively: no pip, setuptools, include/, share/

Expected size reduction:
  Before: 17GB
  After:  ~3GB (82% reduction)

Breakdown:
  - CUDA base: 0.4GB
  - Paddle: 0.7GB
  - PaddleOCR: 0.8GB
  - OpenCV-headless: 0.2GB
  - Other deps: 0.6GB
  Total: ~2.7-3GB

Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>

pyproject.toml

@@ -11,7 +11,7 @@ authors = [
 dependencies = [
     "fastapi==0.128.0",
     "uvicorn[standard]==0.40.0",
-    "opencv-python==4.12.0.88",
+    "opencv-python-headless==4.12.0.88",  # headless: no Qt/FFmpeg GUI, server-only
     "python-multipart==0.0.21",
     "pydantic==2.12.5",
     "pydantic-settings==2.12.0",
@@ -20,7 +20,6 @@ dependencies = [
     "pillow==12.0.0",
     "python-docx==1.2.0",
     "paddleocr==3.4.0",
-    "doclayout-yolo==0.0.4",
     "latex2mathml==3.78.1",
     "paddle==1.2.0",
     "pypandoc==1.16.2",