fix(ckpt): handle missing newer MLM checkpoint args

[caohch-1]

What does this PR do ?

Make MLM checkpoint config loading tolerate older checkpoints whose saved argparse.Namespace lacks newer optional Megatron-LM arguments.

Changelog

• Use legacy-safe defaults when selecting MLA/heterogeneous config classes and when normalizing a missing cp_comm_type.
• Avoid redundant direct reads of same-name TransformerConfig fields, and only map renamed checkpoint arguments when those source arguments are present.
• Add regression coverage for missing newer arguments, coupled uneven-pipeline arguments, and preservation of explicitly set modern arguments.

Validation

• Targeted tests/unit_tests/training/mlm_compat/test_arguments.py tests pass.
• ruff check passes for both changed files.
• ruff format --check passes for both changed files.
• git diff --check passes.

GitHub Actions CI

This is an external contribution, so GitHub Actions CI will require approval from an NVIDIA developer.

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 add or update any necessary documentation? No documentation changes are needed for this compatibility fix.
• Does the PR affect components that are optional to install? No.
  • Reviewer: Does the PR have correct import guards for all optional libraries?

Additional Information

Fixes #2127

[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.