Staged benchmark pipelines: arm the next stage's waiter at DONE, not after the audit

When a multi-stage benchmark (smoke, then session 1, then session 2) runs on a shared host via gap files or locks granted by a reviewer, it's tempting to wait for the reviewer's audit of stage N before arming the waiter for stage N+1. That leaves the host idle for the whole audit (an hour in one case), and the slot goes to whichever other job was armed. Arm the next waiter as soon as the previous stage reports DONE. The waiter only polls for its gap file, so the reviewer can still withhold the gap if the audit fails. Keep 'armed' and 'granted' as separate states.