Set GRADIO_TEMP_DIR=/tmp for file serving

fix: simplify entrypoint — models pre-downloaded in Docker image

fix: re-add SyncNet download (required by MuseTalk)

fix: parallelize model downloads to fit 30-min startup window

fix: add download retry, validate models, use exec for clean process mgmt

fix: download diffusion_pytorch_model.safetensors for SD VAE

fix: pin gdown<5.0 to fix missing gdown.cli module

Fix: pin huggingface_hub<1.0 to fix CLI import error

Fix entrypoint: download model weights before starting app

Fix: download model weights at startup

Create entrypoint.sh (#2)