Support async scheduling for prefill transitions

[lmcafee-nvidia]

Summary

• Run non-chunked prefill and mixed-batch transitions through the async scheduling path.
• Drain a pending async forward before admitting newly ready work, preserving one model forward per controller step.
• Transition surviving prefill requests into decode during resolution while preserving legacy request-row movement.
• Keep chunked prefill outside the async scheduling path.

Architecture

The engine treats admission draining as a generic async scheduling transition. When new work is ready while an async forward is pending, the controller consumes and resolves that pending result without launching a successor forward. The next engine iteration can then admit the waiting request without mixing two forwards into one controller step.

For non-chunked prefill, the controller resolves the completed prompt forward, prepares the surviving requests for decode, and launches the next forward through the same async scheduling machinery used by steady-state decode. Request compaction follows the legacy row-movement order so request-local sampling state remains attached to the same logical requests.

[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.