fix(core/service): defer tracing exporter termination until graceful shutdown completes

[leevec]

Currently, the tracing provider is torn down prior to the completion of the service’s graceful shutdown sequence. Consequently, spans emitted during this phase—such as those from in-flight requests or cleanup routines—are dropped, as the exporter terminates before it can flush pending data to the backend.

[kevwan]

Review

Good fix for a real observability gap — spans emitted during graceful shutdown were being silently dropped.

Problem: The previous code registered trace.StopAgent() as a shutdown listener via proc.AddShutdownListener. This ran during signal handling (SIGTERM/SIGINT), which happens concurrently with or before the server finishes draining in-flight requests. As a result, spans from the draining phase were dropped.

The fix:

1. Removes trace.StopAgent() from proc.AddShutdownListener
2. Adds TearDown() to ServiceConf that calls trace.StopAgent()
3. Calls TearDown() from rest.Server.Stop() and zrpc.RpcServer.Stop() — i.e., after the server has finished draining

Concerns:

1. Missing gateway server — gateway.Server has its own Stop() method. This PR updates rest and zrpc servers but not the gateway. Should gateway.Server.Stop() also call TearDown()?

2. Signal-triggered shutdown (not through Stop()) — If the process is killed via signal and Stop() is never called explicitly (e.g., in tests or unusual shutdown flows), trace.StopAgent() won't be called. The previous behavior ensured it ran on shutdown signal. Consider whether a belt-and-suspenders approach (both shutdown listener + Stop()) with deduplication would be safer.

3. RpcServerConf stores the conf struct by value — This means TearDown() is called on a copy of ServiceConf. Since trace.StopAgent() likely uses global state, this should be fine in practice, but it's worth noting.

4. Tests — The test changes (defer c.TearDown()) are correct. However, tests for rest.Server.Stop() and zrpc.RpcServer.Stop() calling TearDown() would increase confidence.

Overall: The direction is correct and the fix improves trace completeness during graceful shutdown. The gateway concern is worth addressing before merge.