Each GPU peer lane launches, drains and counts on its own stream, so a
submit on a parked lane no longer waits out whichever block is still
working. Measured on an RTX 3070, the timed lane falls from 139.918 ms
to 0.140 ms against a 150 ms feeder, and from 389.871 ms to 0.122 ms
against a 400 ms one.

GpuPeer::wait returns Err for a slot that completed carrying a failed
status, where it previously returned Ok. This is the breaking change
in the release: a caller testing only for an error was reading an
unfilled payload as an answer.

Also adds submit_user_on_lane, a MAX_POLLER_LANES bound, bounded team
barriers on both sides, and FLYNNEL_PROFILE_SAMPLES for inspecting the
host dispatch calibration.