Precompiled wheels with GPU

[danielhers]

There are wheels for CPU on PyPI but not yet for GPU. Need to build on a machine with CUDA and upload, possibly as a separate project dynet-gpu. Related to #657 and #841.

[danielhers]

This is not so simple, as manylinux wheels are currently built on the pypa docker, which is CentOS5 and doesn't support CUDA (note that in our Travis CI configuration, CUDA is installed "natively" on the Linux machine with manual install and not pip). In fact, TF, which does provide GPU-supported wheels, builds them on a newer toolset and so they are not really manylinux1-compatible (tensorflow/tensorflow#5033 and tensorflow/tensorflow#8802). It's possible to do that for DyNet too but without a pre-made Docker image it's complicated.
Maybe when manylinux2 exists, it will be easier.