Run async scheduling primers as standalone controller steps

[lmcafee-nvidia]

Summary

• Return an explicit controller-step result that distinguishes primer-only calls from sampled outputs.
• End an async-scheduling controller call immediately after launching its forward primer.
• Re-enter the controller inside the engine and synchronous wrapper without rescheduling requests or crossing the engine step boundary.

Architecture

Each controller invocation now launches at most one model forward. A missing logits state produces a primer-only result; the caller immediately invokes the controller again to consume those logits and run the normal prepare, sample, forward, and resolve flow. Request scheduling, engine bookkeeping, counters, and externally visible output remain attached to the completed sampled step rather than the internal primer transition.

Performance

DFW benchmark.py on the dense 3.58B model, using one real 3,675-token prompt, OSL 512, batch size 1, and full-iteration CUDA graphs. The table reports three independent recorded trials per branch; each trial followed one warmup run.

Branch	Measured throughput (tok/s)	Mean (tok/s)
Overlap base	104.0, 106.1, 104.7	104.9
Single-forward	105.8, 104.1, 101.6	103.8

The mean delta is -1.1%. The tested single-forward commit predates the preserving rebase but has the same stable patch ID as the current PR commit.

[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.