Fixed a bug in AccuracyLayer.

[ryanflower]

gpu_diff of bottom[0] used as temporary buffer needs to be set back to
zeros. Otherwise the AccurarcyLayer may propagate back wrong gradient
values in training.

[shaibagon]

This is not a good fix: if the prediction layer is shared with other layers (e.g., "SoftmaxWithLoss") resetting diff might reset the gradients coming from the loss layer.

Please see PR #5987 for a fix for this issue.

[ryanflower]

I see your point, though I think split layers are automatically added to avoid such cases.

A better fix is to use bottom[0]->diff().use_count() to check whether the diff memory is shared with other layers. If so, initialize and use an internal buffer; otherwise reuse bottom[0]->diff() and reset it to zeros after usage.

[shaibagon]

@ryanflower this is an interesting point you are making. Can you please move this discussion to the open PR #5987?