support online training

[grahamgower]

We currently simulate and/or sample from the vcfs, then train. This is inefficient when doing CPU simulations and GPU training. We should put simulation results into a shared queue and when it's big enough, drain the queue and pass the contents to the discriminator for training.