feat: Swin Transformer v1 Support

[danielclough]

Swin v1

This PR adds support for Swin Transformer, a hierarchical vision transformer that uses shifted windows for efficient self-attention computation.

Implementation Details

Model (candle-transformers/src/models/swin.rs):

• Full Swin Transformer v1 architecture with hierarchical feature maps
• Window-based multi-head self-attention with relative position bias
• Shifted window mechanism for cross-window connections
• Patch merging for downsampling between stages
• Supports HuggingFace pretrained weights format (separate Q/K/V projections)
• Linear computational complexity with respect to image size

Example (candle-examples/examples/swin/):

• Command-line interface for ImageNet classification
• Support for all official model variants:
  • Swin-Tiny (28M params, 224×224)
  • Swin-Small (50M params, 224×224)
  • Swin-Base (88M params, 224×224 and 384×384)
  • Swin-Large (197M params, 224×224 and 384×384)
• Automatic weight downloading from HuggingFace Hub
• CPU and GPU support

Usage

# Run with default Swin-Tiny model
cargo run --example swin --release -- --image strawberry.jpg

# Run with larger model variant
cargo run --example swin --release -- --image photo.jpg --which large-large

Testing

Tested with all model variants using weights from microsoft/swin-*-patch4-window* on HuggingFace Hub. Implementation matches HuggingFace's weight format for compatibility with official pretrained models.

Notes

• Swin v2 support marked as TODO (not yet implemented)
• Includes comprehensive documentation and README with usage examples