[Refactor] Pass GLM DSA top-k IDs explicitly

[jayhenry]

Summary

• Pass contiguous int32 dsa_topk_ids explicitly through GLM decoder layers and MTP inputs/results, removing the SequenceContext cache and manual cross-layer lifecycle runtime.
• Use one saved-tensor offload window per decoder call for hidden states and the final shared-layer top-k IDs, with prefetch and reserved pinned buffers; MTP top-k offload remains excluded.
• Keep sharing topology stateless and add coverage for explicit top-k propagation, checkpointing, SP, EP, micro-batches, and MTP.

Validation

• Full pre-commit suite: docformatter, pyupgrade, ruff, ruff-format, mypy, and pydantic-extra-check.
• Relevant unit/integration regressions passed for DSA Torch/TileLang/cuDNN backends, SP2, checkpoint, MTP shared logical depth, EP2 micro2, the optimized 8-GPU engine path, and generic non-DSA MoE.
• 16K, 8xGPU AdamW: rank-0 effective TGS 10,893.7 vs 10,957.3 baseline (-0.58%; below the 11K absolute target); peak allocated memory 104.69 vs 105.26 GB.
• 32K, 8xGPU Muon SP2+EP2+micro2: full-rank sequence TGS 1,769.0 vs 1,599.3 baseline (+10.61%); peak allocated memory 101.87 vs 102.29 GB; VmPin 42.87 GiB in both runs.

Follow-ups

• Use selective checkpoint recompute to keep the top-k indexer out of replay recomputation.
• Evaluate fullgraph=True separately.