fix(token-usage): serialize background pauses, close pacing gaps

Review follow-ups on the shared background ledger:

- Pauses are now serialized reservations (paused_until): concurrent
  sweep and flush settlements queue their pauses instead of sleeping
  them overlapped, so combined background CPU holds one duty cycle
  rather than one per thread — and a capped flush sleep forgives
  nothing, since the reservation stands for the next segment.
- flush_metrics charges only its store segment; the inner drain paces
  itself, and wrapping it billed the drain's own pauses as fresh work.
- The paced drain cap counts iterations, not uploads, and the
  filtered-records skip path pays its pause too — a backlog of
  malformed or excluded records no longer loops unpaced to the 30s
  deadline.
- The backlog-drained reconcile runs after notify tasks as well (one
  SELECT when their inline pass left nothing): flags left by a failed
  inline reconcile no longer wait for the next 30-minute sweep tick.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>