non-slow version of inputTensor

[neubig]

Currently the inputTensor function is slow because it swaps the memory layout from numpy's row-major order to DyNet's column-major order. We are thinking of adding inputTensorTranspose that inputs a tensor and transposes its dimensions at the same time. This will allow for inputting numpy tensors without swapping the dimensions, which will probably be important when inputting data in this way.