Added a MirroredStrategy to train models on multiple GPUs if they are available

3 jobs for master in 19 minutes and 16 seconds (queued for 2 seconds)