fix(v2): drain inflight requests

[kolesnikovae]

Currently, because the HTTP/gRPC server is a dependency of services and is shut down last, there is a possibility that it may receive requests after de-initialization. In most cases, if the deployment is managed properly, the issue is barely noticeable: the load balancer's grace period allows clients to remove the peer from the "pool."

However, when dealing with client-side discovery (e.g., via memberlist/ring, Kubernetes, DNS, etc.), nothing prevents clients from sending requests to a target that is already de-initialized but not yet excluded from the list of available/discoverable instances. The fix ensures that a request will either be handled or rejected appropriately.

Note that heartbeats/health checks are not helpful in this case (and in many other scenarios).

The problem is not specific to the experimental modules; however, the fix is only enabled for them.