Raise nicer error if parallel can't divide batch by num devices

[carlini]

If you try and run a parallel module over a batch of examples where the batch shape doesn't divide the number of devices, the code crashes in numpy land. Instead, throw a nicer error message.