Fix: increase generation timeout from 10 to 20 minutes for T4 with CPU offload

Fix: remove invalid --guide_scale arg from generate.py command

Fix: download model at startup via huggingface-cli, remove build-time download

Switch to official Wan2.2 inference code for native T2V+I2V support

Fix: enable CPU offload + attention slicing for T4 (14.75 GB VRAM limit)

Fix: total_mem → total_memory for PyTorch 2.8 compatibility

Upgrade UI: gallery, progress tracking, tabs, image preview, polished design

Add Wan2.2 TI2V 5B FastAPI app: text-to-video + image-to-video