Multi-gpu support for pytorch

[SteffenCzolbe]

The PR adds multi-GPUs support to the pytorch train script. This function is already present in the Tensorflow implementation, but was absen from the pytorch one.

Command-line arguments are unchanged, and in line with the Tensorflow implementation.

New supported functions:

• Train on a single GPU, that isn't the first one. Example: command line arg "--gpu 3" trains only on Cuda device 3
• Train on any number of GPUs. Example: command line arg "--gpu 0,1" trains on GPUs No. 0 and 1. Parallelism is achieved by splitting the batch among the first dimension. If the batch size is less than the number of GPUs, an error message is thrown (behavior in-line with TensorFlow backend).

Verification:

Implementation was verified by benchmarking training speed on synthetic data. An almost linear speedup was achieved when scaling from 1 to 4 GPUs.

[adalca]

Thank you @SteffenCzolbe , we'll take a look at this and do the pull in a bit.