fix: pass include_buffers=True to init_empty_weights for Gemma-4 support (#1169)

### What does this PR do?

Type of change: Bug fix

Pass `include_buffers=True` to `init_empty_weights()` when computing the
device map for HuggingFace models. Gemma-4 registers its model
parameters as buffers rather than parameters, so without this flag they
are not accounted for during device map computation, causing incorrect
placement or OOM errors.

### Usage

```python
# No API change — fix is internal to get_model() in example_utils.py
# Simply run HF PTQ with a Gemma-4 model as usual:
python hf_ptq.py --pyt_ckpt_path google/gemma-4-... --quantize ...
```

### Testing

Manually tested with Gemma-4 model loading via `hf_ptq.py`.

### Before your PR is "*Ready for review*"

Make sure you read and follow [Contributor
guidelines](https://github.com/NVIDIA/Model-Optimizer/blob/main/CONTRIBUTING.md)
and your commits are signed (`git commit -s -S`).

Make sure you read and follow the [Security Best
Practices](https://github.com/NVIDIA/Model-Optimizer/blob/main/SECURITY.md#security-coding-practices-for-contributors)
(e.g. avoiding hardcoded `trust_remote_code=True`, `torch.load(...,
weights_only=False)`, `pickle`, etc.).

- Is this change backward compatible?: ✅
- If you copied code from any other sources or added a new PIP
dependency, did you follow guidance in `CONTRIBUTING.md`: N/A
- Did you write any new necessary tests?: N/A
- Did you update
[Changelog](https://github.com/NVIDIA/Model-Optimizer/blob/main/CHANGELOG.rst)?:
N/A

### Additional Information

Gemma-4 uses buffers instead of parameters for some model weights,
requiring `include_buffers=True` for correct device map estimation with
`accelerate`'s `init_empty_weights`.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

* **Bug Fixes**
* Improved model initialization in quantization examples by ensuring
buffers are properly included during temporary model construction,
resulting in more accurate device mapping inference for model
optimization.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Signed-off-by: James Shen <yueshen@nvidia.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: 2 people

examples/llm_ptq/example_utils.py

@@ -643,7 +643,7 @@ def has_pack_quantized_config(config):
                 auto_model_module = getattr(transformers, architecture)
                 from_config = auto_model_module._from_config
 
-            with init_empty_weights():
+            with init_empty_weights(include_buffers=True):
                 # When computing the device_map, assuming bfloat16 precision by default,
                 # unless specified by the hf_config.
                 torch_dtype = getattr(hf_config, "torch_dtype", torch.bfloat16)