Improve Remove BG with classical edge-aware segmentation

[Vxiey]

Summary

Rebuilds Remove BG as a fully local classical computer-vision pipeline. It uses no AI/neural model, no cloud API, no model download, no OpenCV and no new runtime dependency.

What changed

• Replaces the old single-reference-colour flood fill with adaptive border colour modelling.
• Uses deterministic 1/2/4-colour clustering for plain, uneven and tiled/room backgrounds.
• Filters high-gradient border samples so a subject touching the frame is less likely to become a background prototype.
• Uses chromatic RGB edge barriers, not only brightness edges, to protect skin, clothing and other equal-luma colour boundaries.
• Region growing combines global border-colour confidence with conservative local-gradient continuation for shadows and mild lighting gradients.
• Adds a classical trimap/confidence layer: uncertain pixels remain foreground; only verified border-connected background is removed.
• Adds connected-component safety checks and fails closed if the remaining subject is implausibly small or fragmented.
• Feathers only the immediate cut contour; deep foreground stays opaque and deep background stays transparent.
• Runs segmentation on a bounded proxy (max ~900k px / 1200 px dimension) and projects the final mask back to source resolution.
• Preserves existing alpha and cancellation behavior.
• Emits diagnostic metadata for background clusters/weights/spreads, edge threshold, trimap percentages and largest foreground component.

Regression coverage

• Internal same-colour details that are not border-connected stay opaque.
• Multi-colour/tiled background is removed while a skin-like central subject remains intact.
• Similar-luma but different-colour subject edges are protected.
• Existing transparency remains untouched.
• Uniform images fail closed instead of deleting the whole image.
• Cancellation is honored.
• Tests assert that ONNX Runtime, rembg, OpenCV, TensorFlow and Torch are not required.

The implementation is intentionally conservative: preserving some difficult background is preferred over cutting holes through a face or other foreground detail.