Stage 2 phase 2: add x2t_image, image_edit, video_edit with shared _run_lance_task helper

Write t2i prompt file pretty-printed so dataset loader detects dict format

Stage 2: add text-to-image task with dedicated _image_inference helper

Bump GPU duration to 240s and raise defaults to 480p/50 frames/30 steps

Lower GPU duration to 120s and adopt smaller default resolution

Disable Gradio 6 SSR mode and move theme to launch() for client compat

Pull top-level config.json and confirm Wan2.2_VAE.pth pattern

Also download Qwen2.5-VL-ViT subdir for the vision encoder

Preload libcudart.so.12 via ctypes before transformers loads flash_attn

Import torch first to register CUDA libs before flash_attn loads

Initial Space deploy, ZeroGPU adapter for Lance