[FEA] multiGPU support for metrics

[bschifferer]

Is your feature request related to a problem? Please describe. We use TensorFlow+Keras+Horovod for multi-GPU examples. Some evaluation metrices cannot be averages, for example AUC. As a user, I want to calculate a correct AUC over multiple GPUs.

Additional context I found this for TF1.x https://github.com/horovod/horovod/issues/444 Or https://stackoverflow.com/questions/56552397/custom-metric-using-scikit-learns-aucroc-calculator-with-tf-keras

The 2nd one suggest to implement a custom callback.