Add the centered version of RMSProp
This is based on the RMSProp implementation in TF.Keras: https://github.com/tensorflow/tensorflow/blob/v2.3.1/tensorflow/python/keras/optimizer_v2/rmsprop.py
Having a centered RMSProp is necessary for proper performance of certain algorithms, such as the original DQN: https://github.com/google/dopamine/blob/master/dopamine/agents/dqn/configs/dqn_nature.gin#L23
