Swin upgrade

[TeddyLiang01]

Summary

This PR adds opt-in Swin Transformer (SwinT) backbone support to detectree2 training + inference, with the SwinT_detectron2 implementation vendored under detectree2/third_party so users don’t need to manually clone/install extra repos. When Swin is enabled, detectree2 can merge the Swin config, load/download compatible weights, and run training/inference without ResNet-specific assumptions.

Key changes

New Swin utilities: detectree2/models/backbones_swin.py

• prepare_swint_config(...): makes vendored swint importable, registers Swin config nodes, returns YAML to merge.
• ensure_swint_weights(...): uses provided weights path or auto-downloads default weights to a user cache dir.

Training config support: detectree2/models/train.py

• Adds use_swint_backbone, swint_config_path, swint_weights_path to setup_cfg.
• If Swin is selected: merges Swin config and sets cfg.MODEL.WEIGHTS to Swin weights, otherwise preserves existing model-zoo behavior.

Trainer resume/load compatibility: detectree2/models/train.py

• resume_or_load now skips ResNet stem.conv1 logic when Swin is enabled (no conv1/stem assumptions).

Inference channel handling (Swin-only): detectree2/models/predict.py

• Adds _using_swint_backbone(predictor) and only when Swin is active, forces inputs to 3 channels (repeat 1-band, drop >3, pad <3) to match Swin weights expectations.

Vendored third-party + packaging: setup.py includes vendored YAMLs + swint/*.py in package_data; .gitignore avoids committing .pth files.

Usage

Example training:

cfg = setup_cfg(
    trains=(train_name,),
    tests=(val_name,),
    use_swint_backbone=True,
    swint_config_path=None,          # optional override
    swint_weights_path=None,         # optional override
)

• Omitting swint_weights_path auto-downloads default Swin weights to cache.
• If use_swint_backbone=False, behavior is unchanged.