Add an ate.actor.oom_kills metric for gVisor and microVM #2427

We have no metric for OOM kills. When an actor's sandbox gets OOM-killed:

• The actor stays RUNNING (Actor stays RUNNING after its sandbox dies #2211), so ate.actor.crashes doesn't move. If a later operation fails, the crash is counted under that operation, with a generic atelet error.
• ateom-gvisor logs a warning when a running actor's sandbox dies #2258 added a WARN with ate.sandbox.oom_kills, but it's a log line, it only fires when the sandbox has no processes left, and it's gVisor only.
• The benchmark (benchmarking: add OOM kills to server telemetry #2391) reads cAdvisor's container_oom_events_total instead. That misses a kill when the cgroup is gone before the scrape, and it can't tell which template the actor belongs to.

So operators can't alert on OOMs or see which template needs more memory.