Support async scheduling for MoE and expert parallelism

[lmcafee-nvidia]

Summary

Enable serial and overlap async scheduling for MoE models and expert-parallel execution.

• Inherit standalone primer steps from the single-forward parent and single-bank Mamba state from the Mamba parent.
• Remove the MoE and expert-parallel compatibility rejections from async scheduling.
• Advance the existing expert-parallel work-consensus cadence when a primer-only controller result immediately re-enters the controller.
• Retain the existing restriction on MoE routing replay.

Architecture

This PR contains only the MoE/EP integration that remains after the generic one-forward-per-step behavior was moved into the single-forward parent. A primer still returns from the controller and immediately continues without normal engine bookkeeping. When expert parallelism is active, the engine accounts for that continuation in the same consensus cadence used by ordinary coordinator iterations, keeping all participating ranks on the same forward slot.

The model, controller scheduling flow, and token dispatcher are otherwise unchanged.

[lmcafee-nvidia]

Closing this archived implementation in favor of the active cumulative stack. Its MTP, EP, Mamba admission, primer, prefill, and step-order architecture is consolidated in #26 and #27.