Explicitly differentiate between variable batch sizes and fixed batch sizes

[phildow]

Layer descriptions track an isBatched boolean and if it is true take the batch size from the input data, not the layer, assuming that the batch dimension supports variable batch sizes (indicated by a None in python and a -1 in c++).

But some models may require a fixed batch size for inputs. I don't think our input code for vector layers cares as long as the right amount of data is supplied in a batch, but image inputs can't parse a fixed batch size along the first or last dimension.