Fix: Resolve KeyError and NameError in TorchGeo backbones weight loading

[clementgilli]

Description

This PR fixes two bugs that prevent the initialization of several TorchGeo-based backbones when using pretrained=True.

1. Missing TorchGeo Bands (KeyError)

Files modified: terratorch/models/backbones/torchgeo_resnet.py, terratorch/models/backbones/torchgeo_vit.py and terratorch/models/backbones/torchgeo_swin_satlas.py

• Issue: When loading weights, TorchGeo models often define their Sentinel-2 bands using formats like "B4", "B3", or "B8a". These keys were missing from the look_up_table, causing a KeyError to be raised during get_pretrained_bands().
• Fix: Added the legacy TorchGeo band formats ("B1" to "B9", specifically including the lowercase "B8a" which was causing silent failures) to the look_up_table. I proactively applied this fix to the Swin Satlas backbone as well to prevent identical crashes.

2. Undefined Function (NameError)

File modified: terratorch/models/backbones/torchgeo_vit.py

• Issue: In the load_vit_weights function, the elif weights is not None block calls a remove_keys() function. This function is not defined anywhere in the file or imported, resulting in a NameError: name 'remove_keys' is not defined.
• Fix: Replaced the undefined remove_keys call with the explicit dictionary key deletion logic (for k in ["head.weight", "head.bias"]: del ...) that is already correctly implemented a few lines above in the same file.

Testing

Tested locally. Models such as seco_resnet18_sentinel2_rgb_seco and ssl4eos12_vit_small_patch16_224_sentinel2_all_dino now successfully translate band names, process the checkpoints, and load their pretrained weights without crashing.