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

2 jobs for master in 25 minutes and 19 seconds (queued for 1 second)