Use subprocess for interpreter

[thomasballinger]

This would:

• make supporting virtual environments pretty easy
• make module reloading more reliable (no more not being able to reload some modules, no more having access to the same modules as bpython)
• get rid of the greenlet dependency (requires c compiler) - no more jumping back and forth between of user code and UI code
• make getting threads to work normally easier
• allow a cleaner evented system around a select loop
• make user subprocesses work correctly
• install one version of bpython (python2.6 - 3.4) and bpython works with all python versions
• let bipython reuse more of our code maybe

It would suck because:

• we would have to do all the completion with interprocess communication
• interprocess communication might not be fast enough for the user experience we need - and leads to caching which could get complicated in the best case, and not work in the worst
• Windows compatibility would probably go further out the window
• some code will have to run in the process (import completion, import observer, completion introspection)
• a simple python repl probably isn't enough - we'll still need an InteractiveInterpreter over there
• there are going to be difficulties with version compatibilities - every combination (n by n) of Python versions ought to be tested
• this may involve reading the IPython code and noticing that it's nicer in some ways than ours 😉

[thomasballinger]

I'm not planning to do this for another release at least.

[thomasballinger]

New thoughts on this:

• if we don't use the Jupyter messaging api, this should be done with https://github.com/pexpect/pexpect
• maybe implement a fun forking thing, similar to https://github.com/thomasballinger/rlundo