Implement GPU nllLoss and array transfer ops.

[dan-zheng]

• Implement nllLoss and nllLoss_grad kernels.
• Add GPUArrayFromSeq expression.
  • This is necessary for initializing the "target" argument of type Array[Int] on GPU.
• Add array transfer expressions.
  • Array transfer ops are necessary for transferring non-Tensor arrays, e.g. target labels of type Rep[Array[Int]].
  • Redefine tensor transfer ops in terms of array copy ops.

GPU nllLoss and nllLoss_grad behave as expected (verified in lantern-cudnn-nll-loss). MNIST CNN works using GPU nllLoss.

[dan-zheng]

Ready for review. There are two main changes:

• Implement nllLoss and nllLoss_grad kernels. These are simple.
• Implement array transfer ops (generalizing tensor data transfer ops).
  • The kernels expect all array arguments (including target) to be on the GPU.
  • Array transfer ops allow us to write target.moveToGPU().

MNIST CNN works using GPU nllLoss.