Redundant application of mask

[torms3]

Mask is being applied to both output activation and target label in cost functions. Applying mask to the resulting gradient (https://github.com/seung-lab/znn-release/blob/master/python/train.py#L102) is redundant.

[xiuliren]

@torms3 could you point out the code masking the activation or target label? I think that we only masked the gradient.

[nicholasturner1]

I put that into cost_fn.py (e.g. https://github.com/seung-lab/znn-release/blob/master/python/cost_fn.py#L30) some time ago to start decreasing the occurrences of NAN values. I have some other updates for the specific implementation there (changing mask_dict_vol), but I think this is the better place for it.