Condition the meta initialization for hf_causal_lm on pretrain

[irenedea]

Before this change, there is an error if the pretrained_model_name_or_path does not have model weights (like safetensors, etc.)

After this change, we can pass in a pretrained_model_name_or_path that only has config files if you are not initializing the model with pretrained weights.

Manually tested in interactive via:

huggingface-cli download meta-llama/Meta-Llama-3-8B config.json --local-dir /local-model/

and setting pretrained_model_name_or_path to /local-model