graph: the phases tokenPhasesOf resolved, or null for a graph whose
method runs none (denoising), which then refuses every token phase.
Whether the composition built a Neural Engine operation.
One new token for each of sequences sequences.
One prompt step of sequences sequences of tokens positions each, in
the phase its planner named.
A verify window of positions (depth + 1) per sequence.
The phase operations of one graph, indexed by what the scheduler decides.
The Neural Engine is a unit the scheduler owns: read once, here, which Neural Engine operations the composition built and the row range each takes. A prompt piece of one sequence, a whole chunk or a tail, whose positions one range holds runs on that operation; every other piece runs on the GPU. One scheduler drives one dispatch at a time, so the unit serves one piece at a time.