feat(moe): support clamped SwiGLU expert paths

[FDecaYed]

What

Extend the core ClampedSwiGLU primitive into MoE-owned paths:

• routed expert fallback and token-weighted fusion;
• TE grouped MLP op-fuser using ScaledClampedQGeGLU when available;
• shared-expert overlap fallback;
• current-main FusedSharedExpertMLP grouped op-fuser;
• explicit TE >= 2.17 and operator-availability gates;
• focused construction, version-gating, fallback, and CUDA integration coverage.

Stack

• Depends on Complete clamped SwiGLU across expert paths #5940.
• Temporary review base: pull-request/5940.
• Files changed contains only two MoE production files and their two tests.
• Before Complete clamped SwiGLU across expert paths #5940 merges, retarget and refresh this PR to main to preserve review history.

Provenance

• Reconstructs the MoE clamp work from [dev] [DeepSeek-v4] Part 2: Hash MoE and SwiGLU clamp #4481, [dev] [DeepSeek-v4] Add ClampedSwiGLU to MoE mlp_op_fuser and add force balance to hash routing #5130, and [dev] [DeepSeek-v4] Part 3: MTP support with mHC and new mHC contract #4518.
• Refactors the corresponding implementation captured in [Tracking] DeepSeek-v4 HybridModel merge plan #5795.
• Original implementation by the source PR authors; DSv4 adaptation by @hxbai.
• Feature cutoff: 2026-07-21. GPTModel integration is intentionally excluded.

Testing

• isort --check-only, ruff, Black --check, compileall, and git diff --check passed.
• An independent read-only review closed non-positive/NaN validation and shared-expert op-fuser gaps.
• Focused pytest collection is blocked locally because this host lacks Torch; execution is left to stack CI.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[FDecaYed]

/ok to test 1412381

[FDecaYed]

Closing this stacked boundary because it is absorbed into P5. The complete clamped-SwiGLU feature across primitive, routed/shared, fallback, and Transformer Engine paths will be published by updating #5940. Tracking PR: #5795.