WriteToArrayOp need to support empty input

[JiayiFeng]

In Dynamic RNN, we need to shrink the memory according to the input of the next time step, and the shrank result will be written to the memory array through WriteToArrayOp. However, if the current step is the last one, there will be no such next input and the shrank result will be a tensor holds no memory. To handle this case, the followed WriteToArrayOp should be able to take an empty tensor as input.