Add Swin Transformer Backbone

[AdonaiVera]

Add Swin Transformer Backbone

This PR adds SwinTransformerBackbone, based on the paper Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. It is the first step to support Grounding DINO in keras-hub, where Swin is the main image encoder.

Related Work

This idea has been discussed in previous threads:

• keras-cv issue #2114 — community request to support Grounding DINO, which requires Swin.
• keras-hub issue #2117 — A request to support Swin-UNETR, a model originally designed for 3D medical image segmentation using Swin Transformers for effective feature extraction.

Included in this PR

• SwinTransformerBackbone model
• Core layers (PatchEmbedding, WindowAttention, etc.)
• Unit tests
• (Presets will be added later)

🧪 Current Status

This PR is still a draft. I’m finishing:

• Fixing some issues with float16 and float32
• Preparing a Colab notebook to:
  • Convert pretrained weights from the original repo
  • Compare the outputs with the original model

Let me know if this direction makes sense, or if you have any comments or suggestions. Thanks!