tensorflow all-reduce & NCCL

TF 2.0 since the beta supports all-reduce without a parameter server, and with NCCL backend:
https://www.tensorflow.org/beta/guide/distribute_strategy

we should use this instead of the parameter server for the benchmark