Handle single-matrix inputs correctly

[bgat]

When feeding a single matrix of inputs to cnnff, i.e. running "live data" through a network that has already been trained, the size() command drops the third parameter in its result because the input looks like two-dimensional data rather than a single instance of three-dimensional data. This missing parameter then triggers an error.

Adding postpad() forces the third parameter to a '1' if size() doesn't provide a different value.