TODO

[TODO]
Docstrings in the Numpy style.

Documentation generally.

Make the tests into a proper tests module and examples module

Sampling from the (variational) posterior - this will involve crawling the network, which means that nodes will need pointers to their parents.
[nodes now all have pointers to parents - this should be easy, though perhaps sampler should be a separate module?

Calculation of the Lower bound
Should this be done *within* the node or should an external module do it?

Gaussian Process Prior nodes.

Partially observed nodes (i.e. handle missing data gracefully)

eliminate corner case where user 'hstacks' scalar Gaussians.


[DONE]
Implement (or think about) a constant class, to allow for __rmul__
Renaming of the messages (currently pass_up_ex, pass_up_prec, which is a little misleading...) [Changed to m1, m2. Also in doc.]
Finish the multiplication node to cope with hstacked matrices [there's still a corner case - see above]
Implement the hstack class