feat(moonep): support 128-rank MindSpeed runs

[chaowick]

Summary

• add MindSpeed rank-table generation and multi-node orchestration for 128-rank runs
• add cross-node stage barriers, idle probing, and stronger launcher/preflight validation
• keep predictable projection workspaces in a persistent UDMA registration arena while preserving logical tensor views

Validation

• patch validated in another environment, as reported by the submitter
• git diff --check origin/main...HEAD

Per request, no build or test suite was run again in this workspace. Hardware validation claims remain scoped to the environment where the patch was tested.