fix: simplify mask processing to match reference model - direct white=remove detection

fix(mask): improve mask detection - handle RGB paint directly, detect magenta and white areas; match reference model behavior

fix(mask): ensure proper mask interpretation - selected areas are removed; add detailed logging

fix(mask): improve mask loading to match standard convention (white=remove); add debug logs

feat: invert-mask option in UI and core; add docs (no secrets)

Enable GPU auto-selection (CUDA if available) with DEVICE override; load TorchScript model on selected device. Reference: https://huggingface.co/spaces/LogicGoInfotechSpaces/Hair_stable_new/tree/main

Initial commit