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

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: https://github.com/ZhengPeng7/BiRefNet/issues/226
It would be great if you can integrate this into the inference pipeline.

Start by tracing the repository's inference pipeline to find where edge foreground handling can be integrated, then study the GPU foreground-estimation approach referenced in BiRefNet issue #226. Done means the pipeline includes the post-processing and produces more accurate foreground separation at image edges.

Written by the indexing model from the issue text.