ParallelFor runs a counted loop across a fixed set of threads created up front
and hands out indices from an atomic counter, so the calling thread shares the
work. A thread-start hook runs once per helper thread.
WorkerLanes keeps one job queue per lane, runs each lane on its own thread and
can pause every lane to drain the queues on the calling thread. A thread-start
hook lets a caller run code once per lane (pin it, name it) without the queue
knowing why.
ThreadPinning reads the sibling groups from the kernel topology and pins the
calling thread to a cpu set. It is stateless and per-thread: a caller reads the
groups, picks the cpus for a thread and pins it, so any mapping of threads to
cpus is expressible.
