fix: Don't drop work queued by the last active table in shuffle-queue

[disq]

Summary

The shuffle-queue dispatcher now gets work from a workQueue that pops and checks for idle under one lock, and wakes on Push, Done or context cancel.

[erezrokah]

LGTM, no blockers. The buffered workDone signal cannot lose a Done, and the second Pop catches relations that the last active table queued.

Follow-ups, not blockers:

• The dispatcher wakes only on Done, not on Push. Relations from a long-running table wait until a worker finishes. A Push signal also removes the need for the second Pop.
• isStarted is not necessary now. It can also make dispatch block forever if the queue starts empty.
• Wait does not use ctx. Consider select on ctx.Done().

[disq]

@erezrokah thanks, all three are handled in 940549b. activeWorkSignal is replaced by a workQueue that owns the queue and the in-progress count under one lock:

• Wake on Push: Push wakes the dispatcher, so relations no longer wait for a Done. The second Pop is gone, because Next pops and checks for idle under the same lock.
• isStarted: removed. Next returns nil at once when the queue starts empty.
• ctx: Next selects on ctx.Done() and returns nil when the context is cancelled.

The tests in work_queue_test.go use testing/synctest, so they do not depend on timing.