Log linear, MLP, RNN use invalid default values

[bpopeters]

It makes no sense to specify input=None or output=None in the headers of the various models' functions, as these arguments are in fact required and the models fail if they are unspecified.

Additionally, input and output are not good choices for variable names -- input because it shadows a builtin and output because it is not an accurate description of what is does -- it is actually the gold label, not the output of any model. Personally I'm partial to X and y because they're simple and are closer to the math.

PR coming in 10 minutes.