Fix opencv variant collision and PyTorch 2.6 torch.load

[andriiryzhkov]

Two CI failures in the conversion pipeline, both blocking the NAFNet job.

opencv collision - mmcv pulls opencv-python, albumentations pulls opencv-python-headless; both own the cv2/ namespace and race in the shared venv, leaving a half-written cv2 (no imread) that corrupts every group. Fix: standardize on headless (core requires it, an override-dependencies entry drops opencv-python). We make no GUI calls, and mmcv works with headless.

PyTorch 2.6 torch.load - 2.6 flipped weights_only to True by default, which rejects our checkpoints (they carry non-tensor config/metadata). Fix: pass weights_only=False in the NAFNet/SegNext/BSRGAN converters - trusted files we ship. Also re-enables denoise-nafnet (drops .skip).

Note: if CI caches .venv or the uv cache, bust it once - a cached env still has the corrupted cv2/ and uv sync won't repair an already-present package.

[TurboGit]

Thanks!