Hugging Face flags safetensors unsafe

This is mostly a heuristics-and-labeling problem, not safetensors suddenly becoming dangerous.

• –HF's security scanner is built to catch risky artifacts like pickle-based checkpoints and suspicious imports, but repo warnings can still trip on edge cases or false positives.
• –Safetensors removes arbitrary-code-execution risk from deserialization, but it does not guarantee the file is benign or well-formed.
• –A `.safetensors` file can still be unsafe to load if it abuses parser behavior, oversized metadata, or other denial-of-service vectors.
• –For teams pulling models, the real check is the exact warning text plus the repo contents, commit provenance, and whether any `.bin` or `.pt` files are also present.