runtime: Add an opt-in slotless poll window after targeted wakes
A thread woken by a targeted wake (jl_wakeup_thread — a sticky or pinned
task was delivered specifically to it) may poll without holding a spinner
slot until the sleep threshold elapses, instead of being force-parked by a
saturated spinner cap. Threads that demonstrably receive directed work
otherwise pay a futex round trip per task: in the parked-consumer channel
pattern the hosting thread is denied a slot by persistent idle spinners
and parks between every round.
Polling uncounted is safe — the wake gate's lost-wakeup argument depends
only on counted spinners' exit protocol, and the thread still parks
through the standard publish + fence + re-check sequence — and it is
demand-proportional: only an actual targeted wake grants a window.
Off by default; Base.__init__ enables it when the workstealing scheduler is selected: the window helps
schedulers with per-thread queues (an extra poller is nearly free; the
workstealing backend's cross-pool pingpong drops 2x, to below the partr
baseline) and hurts ones with shared queues (partr's heaps degrade ~50%
under the extra idle polling, measured by same-build A/B).
This commit was written with the assistance of generative AI (Claude).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
