cothread_wakeup()'s vktimer_ts zeroing races vktimer_thread's reload store — wakeup can be silently consumed, delaying all systimers by up to the stale sleep

Summary

The zeroing-then-SIGINT 'latch' only covers the store->nanosleep-entry window in one direction: if the thread's vktimer_ts store lands after the waker's zeroing while SIGINT is delivered during the rescan, the empty handler consumes the signal AND the zeroing is overwritten - the cothread sleeps the stale-long value and every systimer on every cpu is delayed by up to that value (<=1s, bounded by the reload clamp). Transient hard-timer latency spikes under reload pressure; adjacent to DF-5568 (dead CAS in the same protocol). Single-writer handshake fix.