[WIP] Asynchronous cell execution

[davidbrochart]

This PR allows for asynchronous cell execution. I just duplicated (with an async_ prefix) and made asynchronous the run_cell method and all its callers for now.
An option would be to keep only the async version of the methods and provide a user-facing (non-async) execute method that would accept a asynchronous=False parameter. If False, execute would under the hood emulate a blocking call by running async_execute in the event loop with run_until_complete().

[akhmerov]

Out of curiosity, what is the purpose of this await?

[davidbrochart]

It just yields, so that the event loop can run other tasks.

[akhmerov]

I don't quite understand how this works: the part where it makes sense to await is the kernel client events, but we're not actually awaiting on anything there, as far as I can see.

[davidbrochart]

The purpose of this PR is just to be able to execute several notebooks in parallel in an async context. Previously it could only be done using threads, because this while loop is blocking.

[davidbrochart]

@akhmerov I think I see what you mean. In practice it allows for some degree of concurrency, because of this 1s timeout, so other cells can be run every second.

[davidbrochart]

The user-facing execute method calls async_execute until it completes (same behavior as before). async_execute is an asynchronous method that allows executing several notebooks concurrently in an async context. At the lowest level, we poll for messages and block for 0.1 second so that other cells can be run.
I guess that if we truly want async all the way we should have an async version of jupyter_client's get_msg that would return a Future from zmq.asyncio.Socket.poll().

[davidbrochart]

Trigger CI.

[davidbrochart]

@akhmerov although the tests sometimes fail in the CI (due to very slow machines I guess), I think this is ready to be reviewed. Because jupyter_client's get_msg uses zmq.Socket.poll() which is blocking, I had to poll on get_msg in a non-blocking way (no timeout). But once we get get_msg to use zmq.asyncio.Socket.poll() in jupyter_client, we will be able to await get_msg. Nevertheless I think this PR demonstrates that by having async execution methods we can already have concurrency without using threads.

[davidbrochart]

Closing as #10 has been merged.