Fix Google Maven proxy startup race in Copilot setup workflow

[Copilot]

The failing Actions job was not caused by Gradle resolution itself; it failed earlier because the local Google Maven proxy did not become reachable within the readiness window. Proxy startup was blocked on synchronous donor index construction, so the health check could time out before the server ever bound its port.

• Root cause

  • tools/google_maven_proxy.py built MavenMirrorIndex during process startup.
  • In cold-cache runs, donor archive/index work could exceed the workflow’s readiness timeout, causing the setup step to exit before Gradle started.

• Behavior change

  • Start the HTTP server immediately and defer mirror index construction until the first artifact request.
  • Keep the root health endpoint responsive even when the backing index has not been built yet.

• Concurrency and reconfiguration

  • Add generation-aware lazy initialization so concurrent first requests share one index build.
  • Avoid publishing stale initialization results after handler reconfiguration.
  • Ensure waiters are released correctly if initialization fails and a later request retries.

• Regression coverage

  • Add focused tests for:
    • competing first requests
    • retry after initialization failure
    • stale in-flight initialization during reconfiguration

• Illustrative change

  @classmethod
  def _get_mirror_index(cls) -> MavenMirrorIndex:
      while True:
          with cls._mirror_index_condition:
              generation = cls._config_generation
              if cls._mirror_index is not None:
                  return cls._mirror_index
              ...