Add score averaging and convergence early stop

[CedricHwong]

What does this PR do?

Overview:
This PR extends GradNAS gradient scoring with
(1) optional score averaging
(2) early stopping when gradient scores converge. New search config keys are added with defaults: average_scores, score_convergence_tol, score_convergence_patience, and score_convergence_min_updates. The gradient estimation loop now exits early once all hparam score trackers converge. Unit tests cover the averaging math, convergence boundaries, and disabled-convergence behavior.

Usage

  import modelopt.torch.prune as mtp

  # Example: enable/adjust score averaging + convergence early stop
  pruned_model, search_history = mtp.prune(
      model=model,
      mode="gradnas",
      constraints={"flops": "90%"},
      dummy_input=dummy_input,
      config={
          "data_loader": train_loader,
          "loss_func": loss_func,
          # New options:
          "average_scores": True,
          "score_convergence_tol": 1e-3,
          "score_convergence_patience": 5,
          "score_convergence_min_updates": 10,
      },
  )

Testing

• pytest tests/unit/torch/prune/test_gradnas.py
• pre-commit run --all-files

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.
• Is this change backward compatible?: No Default GradNAS behavior now averages gradient scores and may stop early on convergence, which can change score scale and number of batches processed vs. previous behavior.
• Did you write any new necessary tests?: Yes
• Did you add or update any necessary documentation?: No
• Did you update Changelog (https://github.com/NVIDIA/Model-Optimizer/blob/main/CHANGELOG.rst)?: No

Additional Information

N/A

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[kevalmorabia97]

Hi @CedricHwong we dont use GradNAS anymore these days. Can you please share your pruning use case and what improvements you see after this change?

[realAsma]

#734 (comment)

same as @kevalmorabia97 's comments. Could you please elaborate your use case and what improvements you observe?