Livez

GPU-aware liveness probe.

Unlike /healthz (process-alive only), this exercises the GPU with a tiny CUDA sync. A wedged CUDA context (device-side assert) is unrecoverable in PyTorch, so failing liveness lets the kubelet restart the pod — the only path back to a serving worker (see issue #1025).

It deliberately does NOT consult the lifecycle ready flag, so graceful shutdown draining (when the worker marks itself not-ready) never trips a restart. Wire this to the K8s livenessProbe with a tolerant failureThreshold so a momentarily busy GPU is not mistaken for a wedge.

Returns: 200 OK with "ok" body if the GPU can run a kernel (or this is a CPU worker). 503 Service Unavailable with "gpu unhealthy" if wedged.