More introspection of code submitted to complete_request

The kernel [API function for completion requests (`complete_request`)](http://jupyter-client.readthedocs.io/en/latest/messaging.html#completion) does not specify the level of completeness (no pun intended) a client can expect for a certain input. For example, consider the following code:

``` python
from ipywidgets import HTML
from IPython.display import display
h = HTML("

Suppose I hit **** at the end of the fourth line, the API response will not contain any completion matches. But the kernel does actually have enough information to import and load `HTML` and make its instance methods and attributes available. This is how many IDEs work.

I'd like to request this level of completion be added. It would be a non-breaking update with no API ramifications, and it wouldn't necessarily be an imposition on other kernel authors (though it may encourage them to update their kernels, too).

cc: @ellisonbg @fperez @Carreau