[Dev] Support paged stash with device-initiated GroupedTensor experts

[lhb8125]

Summary

• Enable paged stash for the non-op-fuser TEGroupedMLP GroupedTensor path.
• Mark dynamic saved tensors explicitly through the MCore Transformer Engine adapter.
• Allow HybridEP static-budget paged stash with moe_use_grouped_tensor=True without requiring the TE op fuser.
• Restrict the non-op-fuser paged-stash path to fused SwiGLU/QuickGeGLU activation and add focused coverage.

Dependency and merge order

1. Merge [Dev] Support device-initiated grouped linear without TE op fuser #6847, the dev port of [Main] Support device-init grouped linear module too with when not using TE opfuser #6000.
2. Refresh this branch against the updated dev branch.
3. Merge this PR.

This branch currently contains the #6847 prerequisite commits so it remains buildable before that PR lands. Its effective diff will shrink after #6847 is merged, and the branch will then be refreshed before this PR is merged.

Transformer Engine dependencies

• Improve device-init grouped linear module with single grouped weight support  TransformerEngine#3224 for device-initiated GroupedLinear/GroupedTensor.
• Support paged stashing for GroupedLinear activations TransformerEngine#3423 for paged-stash marking of GroupedTensorStorage.

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.
• CI-faithful GB200 single-file run: tests/unit_tests/transformer/moe/test_paged_stashing.py exited 0. The current mcore-ci-dev image predates TE ci(fix): Parse unit tests in merge-queue #3224, so the new grouped-tensor case is dependency-gated with Installed TE GroupedLinear does not expose use_grouped_tensor instead of failing during module construction.

[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.