Add Shake-Shake model

[rshin]

The model hasn't yet been tested empirically, but it's intended to match "Shake-Shake-Batch" of https://arxiv.org/pdf/1705.07485.pdf on CIFAR-10.

Some potential differences:

• tensor2tensor puts smaller weight decay on larger variables
• Input augmentation details likely differ slightly from fb.resnet.torch
• fb.resnet.torch probably doesn't use gradient clipping
• Different input and output flow. Shake-Shake uses 3x3 conv with 32 channels + batch norm, and 8x8 average pooling + FC, respectively.

[lukaszkaiser]

Looks good, thanks!

[lukaszkaiser]

Thanks for the code, we can improve now.