feat: add graceful shutdown to dashboard and document-service

[pavelanni]

From PR #66 review (GLM 5.1, finding #10):

Both demo/batch/dashboard/main.go and demo/batch/document-service/main.go
use http.ListenAndServe which doesn't handle SIGTERM. In Kubernetes this
can cause request drops during pod termination.

Fix: use http.Server.Shutdown() with signal.NotifyContext.

Low priority — K8s sends SIGTERM with a 30s grace period, and these
are short-lived demo services with no long-running connections beyond
the active A2A call (which has its own timeout).