weight updates should be done with beta=1

[drnikolaev]

As per Natalia:
Line 90 of cudnn_batch_norm_layer.cu should be
cudnn::dataType::one

• weight updates should be done with beta=1, similar to how they are done for convolutional layers, so that updates can be potentially accumulated over a few batches.