Make ow ssh produce a pod you can log into, and not leak it if it can't

[vohonen]

Two related failures. On a RunPod account you don't administer, ow ssh gives
you a machine you cannot reach, and then leaves it running.

Authorise the caller's key:

• ow ssh never passed PUBLIC_KEY to create_pod, so it relied entirely on
  the key being registered in the RunPod account settings. On a shared account
  that assumption fails and you get a pod nobody can log into.
• entrypoint.sh already honours PUBLIC_KEY and writes it to
  authorized_keys, so plumbing it through is enough. It travels as an explicit
  public_key= argument rather than inside env, because start_worker
  rebuilds env from os.environ in dev mode and would drop it.
• New --pubkey, defaulting to <key-path>.pub. If no public key is found we
  warn and carry on rather than failing, since relying on account-registered
  keys is the only thing that works for callers without a local public key, and
  entrypoint.sh overwrites authorized_keys rather than appending to it.

Don't leave a pod billing. There were three ways to lose one:

• wait_for_ssh had no try/except around it, so a timeout propagated out
  before terminate().
• bootstrap_remote called sys.exit() on any failing step, which skipped the
  caller's terminate entirely. This one is the nastiest, because the machine is
  reachable and healthy, so nothing looks wrong. A missing unison in the image
  is enough to trigger it, in every mode rather than just --sync. It now raises
  RemoteBootstrapError and lets the caller decide.
• Both paths now go through one _terminate_after_failure helper, which is a
  no-op for --existing connections, since a machine we did not create is not
  ours to terminate.

wait_for_ssh also discarded stderr, so a rejected key was indistinguishable
from a slow boot: BatchMode=yes fails instantly and the loop just retried for
the full 180 s. It now keeps the last SSH error, reports it, and says outright
when the failure looks like authentication rather than boot time.

Verified against live RunPod pods, both halves:

• provisioning connects on the first attempt, where it previously could not
  authenticate at all;
• pointing --pubkey at a key we do not hold the private half of reports
  Permission denied (publickey,password) and says it is authentication rather
  than a slow boot, then terminates the pod on the spot.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com