fix callback to save train interaction (fixes #222 )
fix collecting aux_input in the interactions
use and set numpy seed rather than python random seed for sampling agents when using PopulationGame in distributed training. Using the PopulationGame in distributed training led to each process having a different seed, thus different senders/receivers/losses were sampled and gradients could not sync across processes/devices. Setting a random seed in the init method of the sampler led to the same batch being used for all devices, thus making distributed training useless. Setting a numpy seed fixes both problems.
Description
Minor bug fixes for multi-gpu training: