[Dev] Support device-initiated grouped linear without TE op fuser

[lhb8125]

Summary

• Port [Main] Support device-init grouped linear module too with when not using TE opfuser #6000 from main to dev.
• Enable Transformer Engine's device-initiated GroupedLinear/GroupedTensor path when the TE operation fuser is disabled.
• Preserve device-side expert token metadata to avoid the legacy .tolist() host synchronization.
• Add grouped weight/bias handling, dispatcher padding support, configuration validation, and numerical coverage.

This PR is the prerequisite for #6845 and should be merged first.

Dependency

• Improve device-init grouped linear module with single grouped weight support  TransformerEngine#3224.

Test plan

• CHECK_ONLY=true BASE_REF=dev ... bash tools/autoformat.sh (Black, isort, pylint, and ruff passed).
• python tools/check_copyright.py <changed Python files>.
• python -m py_compile on all changed Python files.
• Full PyTorch/Transformer Engine tests are left to upstream CI because the local lightweight environment does not include PyTorch.

[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.

[lhb8125]

/ok to test cef04c6