Readiness probe.

Returns 200 if the server is ready to accept traffic. Used by Kubernetes to determine if traffic should be routed to this pod.

Readiness has two conditions:

• Lifecycle state, managed by the lifespan handler: ready after startup completes, not ready during shutdown (draining in-flight requests).
• GPU health: a tiny CUDA sync confirms the device can still run kernels. A wedged CUDA context (device-side assert) keeps returning sticky errors from every inference while a process-alive check stays green, so without this the gateway would keep routing to a dead worker.

Returns: 200 OK with "ok" body if ready. 503 Service Unavailable if starting up / shutting down ("not ready") or the GPU context is wedged ("gpu unhealthy").