EC2 GPU dev loop: use a stopped (not terminated) instance; watch for ICE on launch and on start

For iterative GPU work on EC2 (e.g. g6.xlarge / L4), a launch-build-run-terminate loop spent most of each ~11 min cycle on boot + package install (~2 min), full nvcc rebuilds (~2 min) and rerunning unchanged baselines, with only ~3 min on the code under test. A persistent instance that is stopped between iterations keeps the toolchain, build tree and inputs on EBS. Each iteration is then start, rsync the diff, incremental build, run only affected cases, stop; you pay only EBS while stopped. Caveats: RunInstances for g6.xlarge returned InsufficientInstanceCapacity in some us-east-1 AZs (the error text lists AZs that currently have capacity), so loop over subnets in different AZs on first launch. A stopped instance is pinned to its AZ and StartInstances can also fail with InsufficientInstanceCapacity, so retry with backoff, and keep a fallback of creating an AMI/snapshot to relaunch elsewhere. Keep an idle watchdog that runs shutdown -h now (with the instance's shutdown behaviour set to stop, not terminate) so a forgotten box doesn't bill compute.