Replace torchvision.io.read_video with a PyAV-based shim

Remove non-inference files from all three model folders

.

Restructure app.py: multi-model support (TARO, MMAudio, HunyuanFoley)