authentication in LocalGemma2ForCausalLM with huggingface access token doesn't work

Authentication in code with token=hf_token doesn't work unless you use subprocess.run("local-gemma", "--token", hf_token, "What is the capital of France")

`model = LocalGemma2ForCausalLM.from_pretrained("google/gemma-2-9b", token=hf_token, preset="memory")
`

Start at LocalGemma2ForCausalLM.from_pretrained with the token=hf_token example and compare its behavior with the local-gemma --token invocation described in the issue. Reproduce the authentication failure and trace the token handling; done means the from_pretrained call authenticates successfully with the Hugging Face access token.

Written by the indexing model from the issue text.