Add asynchronous prefill scheduling

[lmcafee-nvidia]

Summary

Add a dedicated asynchronous prefill path so dynamic inference can remain under async scheduling while requests move between prefill and decode.

Architecture

• Route async controller iterations through phase orders tailored to their work:
  • prefill: sample or verify the previous output, resolve it, prepare and admit requests, then launch the next forward
  • decode: prepare, sample the previous output, launch the next forward, then resolve the previous output
  • MTP decode: sample and verify drafts, rewind rejected state, prepare, launch the next forward, then resolve
• Make non-chunked request admission part of the async prefill flow, with non-mutating admission probes and the existing CUDA-graph admission rules.
• Keep the first admitted prefill as a primer-only forward while preserving one base-model forward per controller iteration.
• Preserve GPU-resident sampled-token handoff for decode and defer bookkeeping publication until prepared metadata is ready.
• Keep expert-parallel dummy ranks in the same MTP-before-base-forward order as ranks with active work.
• Extend request preparation and resolution to handle prefill-to-decode transitions and mixed survivor/admission batches without returning to the legacy path.

[lmcafee-nvidia]

Superseded by NVIDIA#5939 now that the overlap prerequisite has merged into upstream main.