[roborluo]

Description

The one-sided MoE combine offset previously depended on the current dispatch
payload size. Changes in payload precision or token count could let a later
dispatch overwrite the previous combine payload while a peer was still reading it.

This change:

• Places combine storage at a fixed, aligned workspace offset.
• Checks dispatch bounds before native writes.
• Uses the same offset for direct-to-workspace outputs.

The existing workspace allocation policy is unchanged.

Test Coverage

• CPU layout tests cover changing dispatch layouts and token counts, overlap
  detection, insufficient capacity, incompatible combine layouts, and lazy
  reservation for both CFT and non-CFT workspaces.
• Constructor tests verify that failed reservation does not publish shared
  workspace state.
• A four-rank GPU regression covers mixed dispatch layouts, uneven token counts,
  empty local ranks, eager execution/CUDA graphs, CFT/non-CFT paths,
  low-precision combine, and direct-to-workspace outputs.

Validation on this branch:

• 22 CPU-isolated tests passed; native CUDA dependencies were isolated/mocked.
• Ruff lint/format, Python compile, and diff checks passed.
• The four-rank GPU regression has not been rerun on this base.