Run control channel in separate thread

[SylvainCorlay]

This makes the control channel run in a separate thread from the shell channel, so that control messages can be processed concurrently to shell requests.

• this is now the recomended approach in the Jupyter protocol.

  For a smoother user experience, we recommend running the control channel in a separate thread from the shell channel, so that e.g. shutdown or debug messages can be processed immediately without waiting for a long-running shell message to be finished processing (such
  as an expensive execute request).

• this is required for ipykernel to support the Jupyter Debug Protocol in ipykernel: we need to be able to remove and add breakpoints while code is running (or even more importantly, process messages such as "continue", "step into", which are by nature to be called during execution.

Fixes #447

[minrk]

Nice! How do you handle maintaining the parent_header state? Processing a control request needs to set parent header for its own side effects (only status?), but must not conflict with the parent header set by the now concurrently outstanding shell request.

Make sure to add a test for this, like:

1. shell request produces multiple outputs (e.g. for i in range(5): print(i); sys.stdout.flush(); time.sleep(1))
2. control request begins and ends in the middle of shell request, ideally ensuring it overlaps at least one or two shell output messages
3. verify that all shell output is properly routed and not captured by control, status messages are handled correctly, etc.

[SylvainCorlay]

Nice! How do you handle maintaining the parent_header state?

Yes, I think we need to have two parent_header attributes depending on the channel like in xeus-python.

now concurrently outstanding shell request.

The debugger tests will essentially be of that nature. Since we are working on plugging the debugger (porting code from xeus-python) on top of this PR, we should probably keep this open until the debugger works in practice.

[SylvainCorlay]

Updates:

• I split the control and shell parent headers
• We now expose a shell_streams property with a deprecation warning
• All tests are passing but somehow, the CI hangs forever in the end. Something must not be shutting down properly.

[blink1073]

Nice work, thank you!

[Carreau]

I think this has broken Jupyter_client ci; there are a bunch of errors wrt control channel.

[Carreau]

missing self.