Re-arm wakeup doorbell on multishot poll termination

[vadimskipin]

The per-CPU wakeup doorbell is a single IORING_OP_POLL_ADD_MULTI submitted once at init: wakeThread writes the eventfd, the multishot poll posts a CQE, and io_uring_enter2 wakes. But a multishot poll is not permanent - the kernel ends it (final CQE with IORING_CQE_F_MORE cleared) on CQ overflow, which IORING_FEAT_NODROP does not prevent (NODROP preserves the CQEs, not the poll's arming). The completion handler never checked F_MORE and never re-armed, so a single termination left the doorbell deaf for the life of the processor: wakeThread's eventfd_write produced no CQE, and the CPU only woke via its park timeout (up to maxWaitNs) - a sticky cross-CPU wakeup-latency cliff.

Factor the arming into ProcessorState::enqueueDoorbell, used by both initialize and handleCompletionQueueSlow, which now re-arms when a doorbell CQE arrives with F_MORE cleared (after the CQ is fully drained, outside the for-each-cqe iteration). Unlike enqueueWakeup, a skipped re-arm is unrecoverable, so it retries through a full SQ ring instead of skipping.