KitForma browser background removal, alpha implementation

U2NetP / U-2-Net by Xuebin Qin and contributors.
Upstream: https://github.com/xuebinqin/U-2-Net (Apache License 2.0).
Unmodified ONNX weights distributed by rembg, release v0.0.0:
https://github.com/danielgatis/rembg/releases/download/v0.0.0/u2netp.onnx
SHA256: 309c8469258dda742793dce0ebea8e6dd393174f89934733ecc8b14c76f4ddd8
Upstream rembg MD5: 8e83ca70e441ab06c318d82300c84806
Architecture: https://github.com/xuebinqin/U-2-Net/blob/master/model/u2net.py
The model already outputs sigmoid probabilities. KitForma uses output d0.

Preprocessing and mask processing adapted from rembg by Daniel Gatis (MIT):
https://github.com/danielgatis/rembg/blob/main/rembg/sessions/base.py
https://github.com/danielgatis/rembg/blob/main/rembg/sessions/u2netp.py
KitForma's TypeScript adaptation uses separable Lanczos3, global RGB-max
normalization, ImageNet means/std, min/max soft mask normalization, and preserves
the source RGB and multiplies the source alpha. Constant/invalid masks are rejected.
No alpha matting or second sigmoid is applied. Floating-point resize rounding can
differ from Pillow; the implementation is not presented as pixel-identical rembg.

ONNX Runtime Web 1.30.0 by Microsoft Corporation (MIT).
https://github.com/microsoft/onnxruntime/tree/v1.30.0
The matching prebuilt WASM and ESM glue are redistributed without modification.
Only WASM execution is enabled, with one thread in a dedicated browser worker.
The bundled frontend retains its upstream license notice. Included dependency
notices are in onnxruntime-ThirdPartyNotices.txt.

See the adjacent full license files. No upstream affiliation or endorsement is
implied. This is a general salient-object segmentation model; difficult edges,
transparent objects and multiple subjects can need manual correction.