KernelManager and KernelClient should be two separate objects

[minrk]

Managing a kernel process and communicating with a kernel are two separate tasks.
The KernelManager conflates both of these, and the code gets weird due to the fact that KernelManagers do not always have the right to start/stop/restart the kernel (e.g. --existing cases).

For instance, in the notebook code, only the management parts of the kernel are used, and the communication (channels) should be done differently, in a tornado-appropriate manner.

The shutdown/restart logic is the main place where these two roles collide, because a clean shutdown request goes via the shell channel. One way in which this could be addressed would be to add the Control channel already used in IPython.parallel. That should enable a clean split of the KernelManager from the KernelClient.

Further, KernelClients should also be connected to the KernelManager, which would enable remote signaling/interrupts via the control channel, eliminating the long annoying "Kernel process is either remote or unspecified. Cannot interrupt." messages.

[minrk]

fixed by IPEP 13