fix(tests): dtype-aware reference cast in test_from_save_pretrained_dtype_inference

[Anai-Guo]

What does this PR do?

Fixes #13869.

test_from_save_pretrained_dtype_inference builds its reference output with a blanket model.to(dtype), then compares it against from_pretrained(tmp_path, torch_dtype=dtype). These two casts are not equivalent, so the comparison can diverge spuriously:

• _keep_in_fp32_modules: from_pretrained keeps these modules in fp32 (see load_model_dict_into_meta in model_loading_utils.py), but model.to(dtype) casts them to dtype.
• Non-persistent buffers (e.g. RoPE inv_freq created with an explicit dtype): these are not saved to the checkpoint, so from_pretrained never casts them and they keep the dtype assigned in __init__. model.to(dtype) casts them unconditionally.

Either case makes the in-memory reference differ from the loaded model and the output assertion fails even though save/load is correct.

This PR replaces the blanket cast with one that mirrors the loader's dtype-aware semantics: parameters and persistent buffers are cast to dtype (except _keep_in_fp32_modules, which stay fp32), while non-persistent buffers are left at their __init__ dtype.

Before submitting

• Did you read the contributor guideline?
• Did you make sure to update the documentation with your changes? (test-only change)
• Did you write any new necessary tests? (this fixes an existing test)

Who can review?

@dg845 @DN6 (diagnosis from #13869 / #13862 review)

🤖 Generated with Claude Code