Simplify completer handling by isolating readline-specific logic more. Also, I removed a hack we were using of storing the completer globally inside sys. This was used for the rare case of embedding ipython-inside-ipython, something which is more of a party trick. If we ever want to support that well, we can write a cleaner solution than polluting sys with 'ipcompleter'. But for now, I'm trying to simplify our completer machinery and remove buggy and/or error-prone hacks. Note that the normal use of embedded ipython inside of user's programs isn't affected by this.