fix: Use trust_remote_code_for_model() in all from_pretrained calls

[kendrickb-nvidia]

Summary

• Some AutoConfig, AutoModelForCausalLM, and AutoTokenizer from_pretrained calls were missing trust_remote_code=True for nvidia/ models (e.g. Nemotron), causing ValueError when loading models with custom code
• Consolidates the check into a single trust_remote_code_for_model() in llm/utils.py, called by all 8 ModelMetadata subclasses, populate_derived_fields, LLMPromptConfig.from_tokenizer, and HuggingFaceBackend
• Removes the redundant TrainingBackend._trust_remote_code_for_model() method

Test plan

• Verified nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 loads successfully through the SDK SafeSynthesizer pipeline
• Unit tests pass (make test)

Other notes

Created #231 to followup and make modifying this behavior use configurable.

Made with Cursor

[kendrickb-nvidia]

Also more from_pretrained() like calls during generation that also need to pick up this trust remote code logic. Need to follow up on those before merging.

[binaryaaron]

good consolidation -- the populate_derived_fields fix is the important one and the duplicate removal from TrainingBackend is clean.

one thing: the utils.py change narrows the type from str | Path to str and drops the str() coercion. huggingface_backend.py passes self.params.training.pretrained_model which could be a Path from config and Path doesn't have .startswith().

this will conflict with #286 and #287 since both touch every subclass init in metadata.py.

I'd plan to land #286 first and then either #287 or this. the rebases should be pretty straightforward.

longer term the repeated AutoTokenizer.from_pretrained(..., trust_remote_code=...)/AutoConfig.from_pretrained(...) pattern in every subclass is asking to be a base-class helper. #231 (making it configurable) would be a good time to do that.