TopK for Tensor?

Faster TopK functions for Tensor can be useful for processes like beam search. Correct me if I'm wrong, currently I have not found efficient way to do this directly on GPU in dynet.

It seems that tensorflow and pytorch adopt two different methods for this (but seemingly not with Eigen), which can be good references or resources:
Tensorflow (heap-based):
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/topk_op_gpu.cu.cc
PyTorch (radix-selection):
https://github.com/pytorch/pytorch/blob/master/aten/src/THC/THCTensorTopK.cuh

Will it be ok to adapt them into dynet? (Putting the relevant codes in `third_party` with corresponding comments about lisence and sources.)
Maybe something like:
`pair TensorTools::topk(const Tensor& v, unsigned dim = 0, unsigned k = 1);`
If ok, I would like to work on this.

Start with the proposed TensorTools::topk API and the linked TensorFlow and PyTorch GPU implementations. Determine how the relevant code could fit under third_party with license and source comments, and define the requested dimension and k behavior. Done means DyNet exposes an efficient GPU TopK operation for Tensor.

Written by the indexing model from the issue text.