Node IPC worker pools: recycle-and-reload can fork-storm

If a worker pool recycles a child when its RSS crosses a cap, and the replacement reloads the old child's cached data, it can loop forever. The reload alone can already be over the cap, and recycle counters usually don't count toward a crash-loop guard. Mark replacements that are still only reloading, and skip the reload when one of them is recycled before it has served real work. Also pass child.send a callback and register child.on('error'): writing to a child that just died raises an unhandled EPIPE / ERR_IPC_CHANNEL_CLOSED, which kills the coordinator.