fix(transformer): AUT-1497 support dual RoPE full recompute

[svcnemo-autobot]

Background

Bridge Gemma4-VL full recompute fails on the selected MCore dev pin because a dual-RoPE tuple reaches autograd checkpoint storage. MCore main already carries the equivalent repair from #5324.

What changed

• Split dual RoPE into tensor checkpoint arguments and reconstruct it inside recompute.
• Preserve the non-checkpointed block path.
• Add focused tensor, dual-RoPE, local-only, and global-only coverage.

Details

flowchart LR
  R[dual RoPE tuple] --> S[split checkpoint args]
  S --> C[autograd checkpoint]
  C --> J[reconstruct RoPE]

Loading

This PR is based directly on dev pin 12eed0a5ac3f1ced6a3a4f76727db530bdbf7391 for coupled Bridge validation.

Tested

• H100 full-checkpoint tests: 5 passed
• Black, isort, Ruff, and Pylint
• Coupled Bridge Gemma4-VL rerun pending

[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.

[svcnvidia-nemo-ci]

/ok to test b3b1e2d