Local background removal in SpriteEditor

The default solid-background mode is native C++ color matching with adjustable
tolerance and a four-connected flood fill from the image (or cell) perimeter.
It does not use an AI model. Disconnected non-background particles are kept;
enclosed background-colored details are kept unless exterior-only is disabled.
This mode changes alpha only and does not reconstruct colors at antialiased edges.

Inference is implemented in native C++ using ONNX Runtime 1.31.1 (CPU).
Runtime: Microsoft, MIT license; see ONNX-Runtime.txt and ONNX-Runtime-ThirdPartyNotices.txt.
https://github.com/microsoft/onnxruntime

Optional AI models: U2Net and U2NetP (lightweight), U-2-Net by Xuebin Qin et al., Apache License 2.0.
https://github.com/xuebinqin/U-2-Net
See U2Net-Apache-2.0.txt.
ONNX model distribution: https://github.com/danielgatis/rembg/releases/download/v0.0.0/u2netp.onnx
MD5: 8e83ca70e441ab06c318d82300c84806
SHA256: 309c8469258dda742793dce0ebea8e6dd393174f89934733ecc8b14c76f4ddd8

Full model: https://github.com/danielgatis/rembg/releases/download/v0.0.0/u2net.onnx
MD5: 60024c5c889badc19c04ad937298a77b
SHA256: 8d10d2f3bb75ae3b6d527c77944fc5e7dcd94b29809d47a739a7a728a912b491

Native preprocessing uses RGB input 320x320, ImageNet channel normalization and
min-max normalization of the first prediction. The rembg documentation was used
as a format reference; no Python runtime or rembg package is included.

Only the alpha channel is changed. Existing transparency, source RGB pixels,
image size and frame positions are preserved. The model is general-purpose and
can miss small pixel-art details; review the preview and adjust the cut threshold.
No image is uploaded and no API credentials are required.
