Svdquant huggingface checkpoint export support

[sychen52]

What does this PR do?

Type of change: new feature

Overview:

Usage

cd ./examples/llm_ptq/
python hf_ptq.py \
    --pyt_ckpt_path Qwen/Qwen3-4B \
    --export_path /home/scratch.shiychen_coreai/quantized_models/Qwen3-4B-svdq \
    --qformat nvfp4_awq_svdquant --kv_cache_qformat none --sparsity_fmt dense --calib_size 8

Testing

exported checkpoint and loaded.

Before your PR is "Ready for review"

• Make sure you read and follow Contributor guidelines and your commits are signed.
• Is this change backward compatible?: Yes/No
• Did you write any new necessary tests?: Yes/No
• Did you add or update any necessary documentation?: Yes/No
• Did you update Changelog?: Yes/No

Additional Information

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 74.63%. Comparing base (fe52b2a) to head (2141906).
⚠️ Report is 9 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #754      +/-   ##
==========================================
- Coverage   74.68%   74.63%   -0.06%     
==========================================
  Files         192      192              
  Lines       18950    18995      +45     
==========================================
+ Hits        14153    14177      +24     
- Misses       4797     4818      +21     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[jingyu-ml]

LGTM overall, including the approach for fusing the QKV and FFN layers. The current resmooth + refusion process means the resulting model is not exactly identical to the original, but this appears to be the only viable option at the moment unless we can fuse these layers during calibration...
Thank you for your work!

[jingyu-ml]

Could you add a check to ensure the model is running on a single GPU and not in a distributed setup? I’m not sure that our current SVDQ implementation works correctly with multiple GPUs. We can remove this check later once we verify that SVDQ calibration functions properly in a multi-GPU setting.

[cjluo-nv]

do we need f64?