Sizing disks for stopped (warm) EC2 build/sim boxes: gp3 baseline is slow, instance store is wiped on stop

When keeping EC2 boxes stopped between jobs (instead of terminating), split storage by lifetime. Put anything that must survive a stop (toolchain, build tree, sha-checked inputs, results not yet pulled) on gp3 EBS. Put big scratch outputs on the local NVMe of *d instance types (c6id, m6id, g6 have it), which is much faster than gp3 but is WIPED on every stop, so pull results before stopping. A gp3 volume defaults to 3000 IOPS and 125 MB/s: writing 10 GB of simulation output takes ~80 s at that rate, so raise throughput or IOPS on the data volume, or keep heavy writes on NVMe. A volume created from a snapshot cannot be smaller than the snapshot's size, so size it from the snapshot, not from the files on it. Set instance-initiated shutdown behaviour to stop, and enable termination protection, so an idle watchdog's shutdown -h now stops the box instead of deleting it. Record the disk class (NVMe vs gp3 + its throughput) on every benchmark row, because I/O-bound timings are not comparable across them.