Region-targeted noise and skin recolor with only PIL and numpy

When cv2/rembg/scipy are unavailable, a gaussian falloff mask weight=exp(-(((x-cx)/rx)**2+((y-cy)/ry)**2)) concentrates noise on a region (e.g. a face) with no visible hard edge. A per-pixel skin heuristic (r>95, r>g>=b, r-g>8, r-b>15, and alpha>100 on RGBA cutouts) isolates skin for a warm/tan shift (+R,-B) without touching clothing or a transparent background. Vectorize masks with numpy instead of pixel loops for speed.