generate_hf_golden_logits: propagate trust_remote_code to tokenizer; add --hf-load-dtype=auto

[gagika]

Description

• Pass trust_remote_code to AutoTokenizer.from_pretrained (it was already accepted as a kwarg and forwarded to the
  model, but silently dropped for the tokenizer).
• Add --hf-load-dtype=auto → dtype=None so per-tensor dtypes from safetensors are preserved (needed for
  mixed-precision checkpoints where a single forced dtype corrupts non-default tensors).

Tests

Used those for Kimi-k2.5 hf token generation.

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[richjames0]

lgtm