Fix a BinderServerTransport crash in the rare shutdown-before-start case

[jdcormie]

The following BinderServerTransport#start() code was completely untested and did not in fact work:

if (isShutdown()) {
      setState(TransportState.SHUTDOWN_TERMINATED);
      notifyTerminated();
      releaseExecutors();

The problem is that if the listener does in fact arrive via start() after shutdown, BinderTransport's shutdownInternal() has already set the state to SHUTDOWN_TERMINATED (which is not a valid transition from itself). It has also already scheduled a call to notifyTerminated() and releaseExecutors(). This causes a duplicate call to transportTerminated and releasing the same executor twice.

This PR changes start() to leave changing state and releasing executors to shutdownInternal()'s. notifyTerminated() either runs then (if already started) or from within start() (if not yet started)

Fixes #12439.

[ejona86]

OkHttp/Netty on server-side don't notify shutdown/termination from the calling thread. They instead trigger operations for the reader and the reader calls the methods. For example, shutdownNow() when still starting will close the connection's file descriptor, which the reading thread will see and notify termination, the same way it does if the client closes the connection. I can see why Binder would be different in this regard.