feat(pipeline): carry mHC residual streams between stages

[FDecaYed]

What

Teach the PP/VPP schedules to carry expanded mHC residual streams across real stage-to-stage edges:

• active edges use num_residual_streams times hidden_size;
• first receive, last send, PP=1, non-mHC, and variable-shape paths retain existing behavior;
• interleaved VPP keeps one uniform expanded shape for its active edges;
• MultiModulePipelineCommunicator remains compatible when no single pp_group exists;
• no new global process-group reads are introduced.

Stack

• Depends on feat(transformer): add native manifold hyper-connections #5936 (which itself depends on feat(recompute): coordinate checkpoint-without-output recomputation #5935).
• Temporary review base: pull-request/5936.
• Files changed contains only pipeline schedules and one CPU shape-test module.
• Before feat(transformer): add native manifold hyper-connections #5936 merges, retarget and refresh this PR to the next stable base/main to preserve review history.

Provenance

• Reconstructs the pipeline shape portion of [dev] feat(mHC): Add basic pytorch implementation of manifold hyper connection(mHC). #2943.
• Refactors the same schedule support captured in [Tracking] DeepSeek-v4 HybridModel merge plan #5795 (commit 07c0d08).
• Original implementation by the [dev] feat(mHC): Add basic pytorch implementation of manifold hyper connection(mHC). #2943 authors; DSv4 adaptation by @hxbai.
• Feature cutoff: 2026-07-21. GPTModel wiring is intentionally excluded.

Testing

• isort, Black --check, ruff, pylint (10/10), compileall, and git diff --check passed.
• CPU/single-process shape tests are included; local execution is blocked because this host lacks Torch. Real PP/VPP behavior remains for stack GPU 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 f12259e

[FDecaYed]

Closing this draft until Round 2. Its logical successor is P3 (mHC PP/VPP transport), rebuilt on main after P2 merges. It will not remain a stacked-review PR. Tracking PR: #5795.