ConfusionMatrix:add(pred, target) has logical error when target is a dim1 tensor of length 1

[pengsun]

because these two lines of code don't check the length of target and the self.mat will be wrongly updated if target:numel()==1.

Note that feeding ConfusionMatrix with target being a dim 1 tensor of length 1 is possible because of this modification to nn.ClassNLLCriterion. I didn't submit a patch because I'm not sure whether we should support or disallow this kind of size.