Migration: Switch to Gradio API (merve/fashn-vton-1.5) and updated token

Fix: Final alignment of dependencies and loading logic

Fix: Force safetensors and fp16 variant to resolve missing text_encoder_2 weights

Deployment: Upgrade to modern dependencies for IDM-VTON support

Fix: Remove strict use_safetensors and handle variant fallback for model loading

Fix inference_wrapper.py imports and structure

Initial backend deployment