Data parallel training support

[neworderofjamie]

Pleasingly, this was actually very easy to do! Basically:

• There's a new class of things called 'communicators' which let you get rank, number of ranks etc and perform basic communications - I've made an mpi4py implementation for now as that's what my old code used
• CompiledNetwork does some basic stuff if a communicator is provided:
  • Only building on the 1st rank
  • Waiting on a barrier before loading
  • Doing the NCCL initialisation
• Compiler subdivides batches across ranks if a communicator is provided and turns on the magic NCCL flag so GeNN generates the additional bits of code (NCCL multi-GPU reductions genn#449)
• Metrics like SparseCategoricalAccuracy get passed the communicator and use it to

Other than that, it's all just passing the communicator around and a few places where the 'full' batch size is used rather than the scaled down one e.g. in the EventProp compiler to scale stuff. I've also added a couple of additional examples (at some point I need to tidy the examples up a bit) which demonstrate how you need to change your code to run across multiple GPUs - mostly just splitting the dataset and turning off progress bars etc apart from on the first rank.

[tnowotny]

wow - that looks surprisingly simple and elegant.
I was first confused where the checkpoints will go but I see now that only rank 0 is writing so that's fine.