feat: improve background removal feathering

- replace hard binary thresholding with smooth
  linear alpha mapping to reduce halos around
  foreground edges especially in low
  resolution images
- normalize now uses a quadratic curve instead
  of hard binarization so that gradient info
  from the model output is preserved
- removebackgroundasync applies a configurable
  linear interpolation between the set min and
  max feathering thresholds so you have partial
  transparency at the edges

src/BGR.Console.Tests/Unit/RemovalCommandTests.cs

@@ -72,7 +72,7 @@ public class RemovalCommandTests : IDisposable
       .ReturnsAsync(maskStream);
 
     _imageProcessorMock
-      .Setup(p => p.RemoveBackgroundAsync(image.Data, maskStream))
+      .Setup(p => p.RemoveBackgroundAsync(image.Data, maskStream, It.IsAny<byte>(), It.IsAny<byte>()))
       .ReturnsAsync(outputStream);
 
     var commandContext = new CommandContext(
@@ -91,7 +91,7 @@ public class RemovalCommandTests : IDisposable
     _imageProcessorMock.Verify(p => p.CreateTensorInputAsync(image.Data, model), Times.Once);
     _inferenceRunnerMock.Verify(r => r.Run(model.Bytes, inputTensor), Times.Once);
     _imageProcessorMock.Verify(p => p.GenerateMaskAsync(outputTensor, image.Width, image.Height), Times.Once);
-    _imageProcessorMock.Verify(p => p.RemoveBackgroundAsync(image.Data, maskStream), Times.Once);
+    _imageProcessorMock.Verify(p => p.RemoveBackgroundAsync(image.Data, maskStream, It.IsAny<byte>(), It.IsAny<byte>()), Times.Once);
     _imageProcessorMock.Verify(p => p.SaveImageAsync(outputStream, It.IsAny<string>()), Times.AtLeastOnce);
 
     File.Delete(imagePath);