Shut down kernels in parallel When stopping the notebook server, it currently sends a shutdown request to each kernel and then waits for the process to finish. This can be slow if you have several kernels running. This makes it issues all the shutdown requests before waiting on the processes, so shutdown happens in parallel. KernelManager (and MultiKernelManager) gain three new public API methods to allow this: * request_shutdown (promoted from a private method) * wait_shutdown (refactored out of shutdown_kernel) * cleanup (refactored out of shutdown_kernel)