Adding post-processing for better background-foreground separation in the edge part.

[minh-nguyenhoang]

Dear author, thank you for your dedication to the open-source community. For this repository, while the model performance is already quite great, the intrinsic property of image sampling make this model fail to accurately capture the foreground at the edge. Some other project has already implement some kind of foreground estimation algorithm, with a really fast one that run on GPU here: ZhengPeng7/BiRefNet#226
It would be great if you can integrate this into the inference pipeline.

[chonk-lain]

@minh-nguyenhoang
thanks a lot for bringing this to my attention. I will look into it and try to integrate that into the library.
The upcoming release of nobg should also include some speedups so this came in the right time :)