perf: switch to FP16 weights — halve download sizes with zero accuracy loss

feat: auto re-detect on model switch when image is loaded

fix: clear previous detections when selecting new image

feat: add model selector for all scales (n/s/m/l/x)

feat: add multi-size model support (n/s/m/l/x) with model selector UI

feat: deploy Next.js web app (replaces basic index.html)

fix: use proper HSLA for semi-transparent bbox fill

Upload folder using huggingface_hub

initial commit