feat: Implement texture caching system with dirty flag optimization (closes #144)

- Add cache_subtree property on Frame for opt-in RenderTexture caching
- Add PyTexture::from_rendered() factory for runtime texture creation
- Add snapshot= parameter to Sprite for creating sprites from Frame content
- Implement content_dirty vs composite_dirty distinction:
  - markContentDirty(): content changed, invalidate self and ancestors
  - markCompositeDirty(): position changed, ancestors need recomposite only
- Update all UIDrawable position setters to use markCompositeDirty()
- Add quick exit workaround for cleanup segfaults

Benchmark: deep_nesting_cached is 3.7x faster (0.09ms vs 0.35ms)

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

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

src/UICircle.cpp

@@ -180,11 +180,11 @@ bool UICircle::setProperty(const std::string& name, float value) {
         return true;
     } else if (name == "x") {
         position.x = value;
-        markDirty();  // #144 - Propagate to parent for texture caching
+        markCompositeDirty();  // #144 - Position change, texture still valid
         return true;
     } else if (name == "y") {
         position.y = value;
-        markDirty();  // #144 - Propagate to parent for texture caching
+        markCompositeDirty();  // #144 - Position change, texture still valid
         return true;
     }
     return false;