Define and use tflex.device; Make peak_gpu_mem_op a no-op on TPUs

dnnlib/tflib/autosummary.py

dnnlib/tflib/network.py

dnnlib/tflib/optimizer.py

metrics/frechet_inception_distance.py

metrics/inception_score.py

metrics/linear_separability.py

metrics/perceptual_path_length.py

metrics/precision_recall.py

tflex.py

training/dataset.py

training/training_loop.py