Add 3D convolution layers and corresponding models

[lmjohns3]

Some operations are going to require 4D inputs and outputs (e.g., 3D convolutions). We should add models to support this, probably just a copy of the theanets/recurrent.py module ported to an additional dimension.