server/grpc: improve graceful stop behavior

[jejefferson]

Summary

This changes the gRPC server shutdown path to behave more like a real graceful stop for long-lived RPCs and streams.

Current behavior in server/grpc/grpc.go is:

1. wait on internal wg
2. only then call grpc.Server.GracefulStop()
3. force Stop() after 1s

That makes graceful shutdown difficult for services with long-running gRPC calls.

What this patch changes

• starts grpc.Server.GracefulStop() immediately when shutdown begins
• adds a configurable graceful timeout via GracefulStopTimeout(time.Duration)
• only force-calls Stop() after that timeout
• waits on the internal waitgroup after transport shutdown has started, not before

Why

For gRPC graceful shutdown, the expected behavior is:

• stop accepting new RPCs as soon as shutdown begins
• allow already-running RPCs/streams to continue
• only force-stop after an explicit timeout

With the previous order, GracefulStop() starts too late and the hardcoded 1s fallback is too aggressive for long-lived calls.

Compatibility

• default behavior is preserved if GracefulStopTimeout(...) is not configured
• this is a minimal transport-level change; it does not change service lifecycle APIs

Validation

I verified the package still compiles with:

go test ./server/grpc -run '^$'