Route xformers memory_efficient_attention through SDPA on Blackwell

Apply appendix bundle: build diff_gaussian_rasterization + nvdiffrast on GPU, preload libcudart.so.13

Drop unused gradio_litmodel3d (caps gradio<5)

Support new ZeroGPU (Blackwell) hardware: bump torch, swap to xformers, build nvdiffrast at runtime

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

fix: Update app

clean code