Initialisation shapes are inconsistent

[dsteinberg]

• we have swapped n_in, n_out in initialisers.py, our 2-d weights are typically (n_out, n_in)
• we need to generalize these functions to more than 2-dim matrices, see the conv layers

There may be a better design here that keeps track of these dimensions... i.e. functions like sample_W in layers.py are horrible!

[dsteinberg]

This also should fix the autonorm initialization, see d25947e