fix(fsdp2): guard uninitialized accumulated grads

[akoumpa]

What does this PR do ?

Adds a common FSDP2 guard so post-backward skips params whose lazy unsharded tensor was never created.

This fixes checkpoint robustness crashes for text-only Ministral3 runs that load through the VLM implementation and leave modality towers unused during backward.

Changelog

• Guard FSDPParam.to_accumulated_grad_if_needed() when _unsharded_param is absent.
• Add a focused unit test for the guard behavior and idempotence.

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? N/A

Additional Information

Short stack trace segment from the failing nemo-ci job:

/usr/local/lib/python3.12/dist-packages/torch/distributed/fsdp/_fully_shard/_fsdp_param_group.py:504: in post_backward
    fsdp_param.to_accumulated_grad_if_needed()
/usr/local/lib/python3.12/dist-packages/torch/distributed/fsdp/_fully_shard/_fsdp_param.py:600: in to_accumulated_grad_if_needed
    or self._unsharded_param.grad is None
E   AttributeError: 'FSDPParam' object has no attribute '_unsharded_param'

Verification:

• pytest -q tests/unit_tests/distributed/test_parallelizer.py::test_fsdp_accumulated_grad_guard_skips_uninitialized_params
• Scoped nemo-ci passed: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/jobs/346201724
• Generated pipeline: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/pipelines/55591672

[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.

[akoumpa]

this modifies global behavior that downstream libraries may assume, so I'm not 100% sure will move forward with it.

[akoumpa]

/ok to test ced1acc