Less fragile interactions with the kernel

[cphyc]

Currently, interactions with the kernel (e.g. to get the list of available functions, or to know which types are compatible which each other) is done by manually executing Python code that prints its content in stdout and listen to IOPub on JLab's side to capture the output.

This is very fragile and error-prone, as any extra output on stdout would add garbage to the data. In addition, errors are not correctly handled and translate into a javascript error (that fails parsing the output to anything sensible).

It would be much better to have a more robust approach.