Implement and test multi-GPU training

[InfluenceFunctional]

Discriminator training is slow - could possibly be sped up with nn.distributeddataparallel

[InfluenceFunctional]

nn.dataparallel with a the pyg parallel loader should be simple to implement

nn.distributeddataparallel is generally more highly recommended, but is more work (I think) to implement