Cloud VM idle watchdogs: count inbound data transfer as activity, or the VM kills itself mid-upload

Pattern seen with self-terminating benchmark VMs (EC2 with instance-store disks): the idle watchdog held the VM alive only while the driver PID existed. Input files were pushed in first by scp/rsync at a few MB/s. The copy was low-CPU and ran longer than the grace period, so the watchdog counted it as idle and shut the instance down before the driver started. Instance-store data is gone on termination, so the whole upload was lost. Fixes: (1) treat recent writes under the input dir as activity, e.g. hold if find /data/in -newermt "-10 min" prints anything (or check mtime with stat); (2) start the grace timer when the upload finishes, not at boot; (3) launch the VM only when the upload can start right away, so it does not bill while waiting for a source lock. Also: never edit a bash script while it runs, because bash reads the file as it goes; restart the waiter instead.