Dev branch for reservoir model training features

[AnnaKwa]

• specify number of synchronization timesteps, rather than batches. i.e. the training config parameter n_batches_burnis replaced by n_timesteps_synchronize
• can predict output variables that are different from inputs (ex. reservoir takes SST state as input but readout predicts an update tendency)
• separate normalization transforms are fit for each set of input, output, and hybrid variables (since in the SST model, the hybrid variables have different scales than the input)