Add C-RADIO as a teacher

[abtonmoy]

Summary

Adds C-RADIO (nvidia/C-RADIOv3-H) as a distillation teacher.

C-RADIO is an agglomerative (multi-teacher-distillation) VFM whose weights are under the NVIDIA Open Model License, which permits commercial use — so it works as a default teacher. It fits the existing setup directly: translator heads are sized from MODEL_FEATURE_SIZES, so it slots in with no architecture changes.

Changes

• foundation_models/vision_models/cradio.py — get_cradio_model / get_cradio_feature, mirroring the existing teacher modules. Input conditioning is handled internally by C-RADIO (pixels passed in [0, 1]); a forward pass returns (summary, spatial_features), reshaped to (B, C, H, W).
• foundation_models/__init__.py — export the two functions.
• preprocessing/feature_extraction_core/models.py — dispatch branches in get_model and get_feature_outputs.
• foundation_models/common.py — add nvidia/C-RADIOv3-H to MODELS and MODEL_FEATURE_SIZES ((1280, 14, 14)).
• configs/training/target_models/cradio.yaml — C-RADIO as a single teacher.
• configs/training/target_models/cddsvr.yaml — the existing cddsv full set plus C-RADIO.

Testing

Verified locally on a single GPU:

• Loads via AutoModel.from_pretrained("nvidia/C-RADIOv3-H", trust_remote_code=True); spatial features are (1280, 14, 14) at 224 input, matching the registered size.
• A DeiT-tiny student distills C-RADIO both on its own and jointly alongside the existing teachers (DINOv2 / CLIP / ViT), with the per-teacher losses decreasing and cosine similarity to the teacher features rising.

Notes

• Feature-normalization stats (mean/var) for the new teacher are generated from the dataset with the existing scripts/preprocessing/calc_feature_mean.py, as documented for the other teachers. Happy to commit the .npy files if you'd prefer them in-tree.
• The checkpoint (nvidia/C-RADIOv3-H) and config naming are easy to change if you'd prefer a different variant or convention.
• C-RADIO's trust_remote_code path pulls in timm, einops, and open_clip_torch; I can add these to the dependencies if you'd like.