Ensure e_score_correction_bias dtype of DeepSeek-V3/R1 is FP32

[xin3he]

What does this PR do?

Fixes # (issue)

Hi, we find that below loading method cannot keep the dtype of e_score_correction_bias in safetensors, so raised this PR to fix it.

Please let me know if this is not the expected change. Thanks.

intel/auto-round#845

from transformers import AutoModelForCausalLM
model_name = "deepseek-ai/DeepSeek-V3.1-Terminus"

model = AutoModelForCausalLM.from_pretrained(
        model_name,torch_dtype="auto"
)

print(model.model.layers[3].mlp.gate.weight) # BF16
print(model.model.layers[3].mlp.gate.e_score_correction_bias) # FP32

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@Cyrilvallez @ArthurZucker

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: deepseek_v3, dots1, glm4_moe, glm4v_moe

[ArthurZucker]

Perfect TY!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.