refactor: avoid requiring Megatron for model_utils import

[alexbowe]

What does this PR do ?

Defers Megatron imports in nemo_rl.distributed.model_utils, so non-Megatron import paths can load the module without requiring the Megatron stack.

This is useful because model_utils is a broad utility module: importing it should not force every environment, test, or non-Megatron workflow to have Megatron available.

Changes:

• Move Megatron imports out of model_utils module scope.
• Keep the GPTModel type annotation available to type checkers without importing Megatron at runtime.
• Add a subprocess regression test that blocks Megatron imports while importing model_utils.

Issues

N/A

Usage

No user-facing API change.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? No documentation update needed; this is an internal import-boundary cleanup.

Additional Information

Validation:

• PYTHONPYCACHEPREFIX=/tmp/codex-pycache python3 -m py_compile nemo_rl/distributed/model_utils.py tests/unit/distributed/test_model_utils_imports.py
• git diff --check
• AIHub: PYTHONPATH=. python -m pytest --confcutdir=tests/unit/distributed tests/unit/distributed/test_model_utils_imports.py -q passed: 1 passed in 1.52s

Claude adversarial review found no blockers.

[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.