tf.scatter_nd in DyNet

Hi,

I want to implement [a pointer-generator network](https://nlp.stanford.edu/pubs/see2017get.pdf). In this [implementation](https://github.com/abisee/pointer-generator) with TensorFlow, it uses [tf.scatter_nd](https://www.tensorflow.org/api_docs/python/tf/scatter_nd) for calculating probability distributions of words.
Can I reproduce ``tf.scatter_nd`` of TensorFlow with the existing operations of DyNet? I think this [input](http://dynet.readthedocs.io/en/latest/operations.html#_CPPv3N5dynet5inputER16ComputationGraphRK3DimRKNSt6vectorIjEERKNSt6vectorIfEEfP6Device) operation, i.e. SparseInputNode, of DyNet seems to have the similar forward calculation with ``tf.scatter_nd``, but it does not calculate gradients of the given values.

Thank you for your help.

Compare TensorFlow's tf.scatter_nd behavior with DyNet's input operation and SparseInputNode, starting from the linked DyNet operations documentation. Check whether the existing operation supports gradients for the supplied values, using the pointer-generator implementation and paper as context. Done means establishing whether existing operations are sufficient or what support would be needed.

Written by the indexing model from the issue text.