Serve the 5-fold CNN ensemble instead of one checkpoint

Serve the 5-fold CNN ensemble instead of one checkpoint

Fix forced-grade display bug found by manually driving the demo

Fix a cp -R staging bug that left the Grad-CAM fix undeployed; remove stray nested src/src and egg-info

Sync the real Grad-CAM regression-head bug fix from the main repo

Force CPU device explicitly (ZeroGPU makes cuda.is_available() report True outside a real grant, breaking auto-detection)

Deploy DR screening Gradio demo (app.py fetches weights from the model repo)